<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by The OpenJaw Data Science Team on Medium]]></title>
        <description><![CDATA[Stories by The OpenJaw Data Science Team on Medium]]></description>
        <link>https://medium.com/@brian_lewis?source=rss-7bd43b8b9097------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/2*keFB2zfHzLNk9KkNakg5qA.png</url>
            <title>Stories by The OpenJaw Data Science Team on Medium</title>
            <link>https://medium.com/@brian_lewis?source=rss-7bd43b8b9097------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 09:42:44 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@brian_lewis/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Identity resolution for travel retailing (part 3)]]></title>
            <link>https://medium.com/the-openjaw-data-science-blog/identity-resolution-for-travel-retailing-part-3-cfc0ddfd2e87?source=rss-7bd43b8b9097------2</link>
            <guid isPermaLink="false">https://medium.com/p/cfc0ddfd2e87</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[identity-management]]></category>
            <category><![CDATA[customer-analytics]]></category>
            <dc:creator><![CDATA[The OpenJaw Data Science Team]]></dc:creator>
            <pubDate>Fri, 13 Dec 2019 14:19:45 GMT</pubDate>
            <atom:updated>2019-12-13T14:19:45.552Z</atom:updated>
            <content:encoded><![CDATA[<p>In this article, we share how we create a single-customer-view at OpenJaw using a new identity resolution algorithm we have developed which is optimised for the travel domain. This is the third and final part of a three-part article on this topic.</p><p><em>By </em><a href="https://www.linkedin.com/in/aurenferguson/"><em>Auren Ferguson</em></a><em>, </em><a href="https://www.linkedin.com/in/john-carney-40328920/"><em>John Carney</em></a><em>, </em><a href="https://www.linkedin.com/in/vignesh-mohan94/"><em>Vignesh Mohan</em></a><em>, </em><a href="https://www.linkedin.com/in/gavin-kiernan-365a359/"><em>Gavin Kiernan</em></a><em>, </em><a href="https://www.linkedin.com/in/charles-wang-174970144/"><em>Yuxiao Wang</em></a><em> and </em><a href="https://www.linkedin.com/in/dr-beibei-flynn-cta-2681129/"><em>Beibei Flynn</em></a><em>.</em></p><h3>Introduction</h3><p>In <a href="https://medium.com/the-openjaw-data-science-blog/identity-resolution-for-travel-retailing-part-2-3b5752cd9930">part 2 </a>of this article described some of the details of the algorithm we have developed at OpenJaw to solve this problem, spanning data cleansing, reducing computational overhead with ‘blocking’ and measuring string similarities with ‘fuzzy matching’. In this, the final part of the article, we will describe how we use a combination of empirical matching rules, machine learning and graph theory to complete the identity resolution process and create a single-customer-view. We also describe the results of experiments that compare the performance of the algorithm to the traditional ‘exact matching’ method.</p><p><strong>Matching entities</strong></p><p>Once we have measured the similarities between fields for all blocks in the data, the next step is to decide if each candidate match does or doesn’t belong to the same entity. Note that it is typical to fuzzy match 5–10 fields for each candidate match. There are 2 commonly used methods to achieve this:</p><ol><li>Create a set of empirical matching rules, or;</li><li>Use machine learning to automatically decide if records match.</li></ol><p><em>Empirical matching rules</em></p><p>The rules based method involves making a series of decisions about the data. Firstly, for each fuzzy matched field, a threshold for a positive match must be picked, e.g. for surname, if fuzzy match score is greater than 0.85 then they are considered the same surname. The threshold values to pick depends on the data sources and regions from where it’s from. Let’s say the data is from a country where generally surnames are very long, a matching threshold of 0.8 may suffice, but in a country like China, where names are typically 2 or 3 characters, an exact match (score of 1) may be required.</p><p>From here, a decision has to be made to decide if candidate matches belong to the same entity or not. To do this we use a hierarchy of rules, starting off with the strongest rule (the one with the least likelihood of being incorrect) and getting less strict until a minimum acceptable criteria is reached. While this sounds somewhat complicated, the pseudocode below gives an example of rules that could be implemented.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VkQFg1-kN4Jp3HmJ6xSRbA.png" /></figure><p>In the above code, the rules commence with the strongest rule. The motivation for it being that for a candidate match, if either passport or id number are both present and if they match, we would have strong confidence that they belong to the same entity.</p><p>The next level of the hierarchy is if given name, surname, date of birth, phone and email (fuzzy) match, they are considered the same entity. The logic for this being if these records have similar values for all these fields, then they are more likely a match than not, on the balance of probability.</p><p>The next rule is similar but only requires one of phone or email to match. While it’s less strict, it provides more flexibility and allows for variability for when data is collected (sometimes not all information will be collected). The minimum criteria for a positive match is if name, surname and date of birth match. It is up to the user to where to draw the line for which rules are acceptable because if they are too relaxed, the system will start incorrectly matching different peoples records together and incorrectly merge different people into one entity.</p><p><em>Machine learning</em></p><p>The other method of deciding if candidate matches belong to the same entity is to use machine learning. This has some major advantages over the empirical rules but also some disadvantages.</p><p>To use machine learning, a labeled dataset, that is a sample of the full ‘candidate matches’ dataset must be created that classifies a list of candidate matches to be positive or negative. This can be very labour intensive as it requires lots of manual (human) work to making these decisions. However, there are methods such as <a href="https://en.wikipedia.org/wiki/Active_learning_(machine_learning)">active learning</a> that can reduce this. Also, the modelling of outcome can be difficult since the majority of the candidate matches in a block won’t match, making this a class imbalance problem.</p><p>The major advantage of using machine learning is that it lets the data and model make the decisions, i.e. there is no need to make decisions about thresholds for each field as was required for the rules based method since the raw fuzzy matching scores are input into the model. The model can also make much more complicated and flexible decisions than any user could hardcoding rules, meaning more positive matches should be captured, thus improving the process.</p><p>Our general view at OpenJaw is to use rules based methods first, and then iteratively improve the identity resolution process later on with machine learning.</p><p><strong>Creating the single-customer-view</strong></p><p>At this stage, we have established which records match on an individual level but these matches are in isolation. Our aim is to create a single-customer-view, at at this stage in the process, only matches between disparate pairs have been established, i.e. record 1 and record 2 belong to the same entity but what about other records that also belong to that entity? How can these records be linked together to create a customer profile? This step of the process is often referred to as <a href="https://en.wikipedia.org/wiki/Record_linkage">record linkage</a>. There are two commonly used methods to link records together:</p><ol><li>Use recursion to iteratively find links that correspond to entities;</li><li>Use graph algorithms.</li></ol><p>We don’t recommend the first method for the travel domain, as it’s fraught with danger. It’s complicated to code and easily susceptible to bugs. The only reason we would recommend this method is if there are no or very few complicated interactions between matches (not true in travel) or if computing resources are scarce.</p><p>The natural choice for a problem like this is to use a graph to link matches together. Graphs are mathematical structures used to model pairwise relations between objects. They are made up of vertices (or nodes) which are connected by edges (links). There are 2 main types of graphs: directed and undirected. Undirected graphs are when edges link two vertices symmetrically. An example of this is Facebook friends; you are both friends with each other, whereas for directed graphs, edges link nodes asymmetrically, i.e. there is directionality component. Twitter is a good example, you may follow someone but they don’t necessarily follow you back.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1023/0*e4MPQbXWvlMwrlV1" /><figcaption><em>Figure 1: Examples of directed and undirected graphs. Facebook ‘friends’ are considered undirected while Twitter ‘follows’ are directional.</em></figcaption></figure><p>The example shows three people: Mary, Jack and Tom. On Facebook Mary is friends with Jack and Tom and since the relationship is undirected Jack and Tom are also friends with Mary. However, on Twitter, Mary follows Jack and Tom but only Jack follows her back, the relationship is said to be directional.</p><p>For identity resolution, we use undirected graphs since there is no directionality between nodes, e.g. if record 1 and record 2 belong to the same entity, record 2 and record 1 also must belong to the same entity by the transitive property. The nodes for our identity graph are the standardised and cleaned dataset and the edges are the matches previously calculated.</p><p>There are a variety of graph algorithms that could be used to find and link records together but we used the connected components algorithm. This algorithm finds sets of connected nodes where each node is reachable from any other node in the same set. A connected component is a maximal connected subgraph of the main graph. That is to say, it will find and link all available vertices, i.e. uses the matches to link together all nodes belonging to an entity. Each subgraph is called a component and a vertex with no edges is also considered a component. In this case, each component is essentially a collection every record belonging to an entity. A unique number, called the ‘OpenJawID’ is assigned to each component. The identities in this data have now been successfully resolved.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*z_pL2KD16kleLDSy" /><figcaption><em>Figure 2: Identity graph showing connected subgraphs circled in red where all nodes inside each red circle are considered belonging to the same entity</em></figcaption></figure><p>So far we have described how to identity resolve data from a variety of sources. The obvious problem being that this data is both static and historic in nature. The whole point of implementing a solution such as this is to get real-time or as close to real-time customer behaviours. Therefore, we need a method to ingest new data and link it to the historical data, as well as assigning previously unseen entities new OpenJawID’s.</p><p>In our identity resolution system, we have implemented a delta process that is automated. Currently, it’s scheduled to run every day and a specified time. It ingests new data from all sources, cleans and standardises it, comparing it to the historically resolved data using the same process; blocking, fuzzy matching and rules/machine learning. If matches are found between the new and resolved data, the new data will receive the same OpenJawID, effectively expanding their profile.</p><p>Records that couldn’t be matched are assumed to belong to new customers. For these records, the original identity resolution process is run because there can be cases where new customers have made multiple transactions in the delta window and these should be linked. For all new customers a OpenJawID is assigned which increments from the previous maximum id number. The delta data (now resolved) is then appended to the historical data to expand the resolved customer database and is ready for the next delta. This process is fully automated and uses <a href="https://airflow.apache.org//">Apache Airflow</a> to manage scheduling and data pipelining.</p><p><strong>How well does the algorithm perform?</strong></p><p>Up until now, this blog series has been conceptual in nature, dealing with the problem of identity resolution, how it works and some of the challenges associated with implementing it. In this section, we will describe a practical example of the algorithm in action using an example dataset.</p><p>Real personal data can’t be used in this article due to it’s sensitive nature. However, it is possible to use publicly available simulated data. While the simulated data has some shortcomings, it is more than enough to demonstrate the power of an identity resolution system like the one we have developed at OpenJaw, since the data includes typos, missing fields and other common data quality challenges associated with this kind of data.</p><p>The simulated data used comes from the <a href="http://users.cecs.anu.edu.au/~Peter.Christen/Febrl/febrl-0.3/febrldoc-0.3/manual.html">FEBRL</a> (Freely Extensible Biomedical Record Linkage) program. For identity resolution we use OpenJaw’s Blueprint product, available on AWS Marketplace in Jan 2020. This is a generic implementation of our identity resolution algorithm, that can be configured by the user using a GUI as shown in figure 3.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*oIc8cchprD84Ng8i" /><figcaption><em>Figure 3. Screenshot of the Blueprint: Identity Resolution user interface.</em></figcaption></figure><p>A sample of the dataset used for these experiments is shown in table 1. The full dataset has 5,000 records, corresponding to 6538 positive matches out of a possible 12.5 million. Some modifications were made to the data to make it closer to the type of data we typically see in the travel industry. Usually, we only see about 70% of records have some sort of ID number, such as passport, national ID or driver’s license. This could be due to agency or hotel bookings where they aren’t required. To mimic this behaviour, approx 30% of the field id_number were randomly removed.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*9UJuBAjPoQz8fPjP" /><figcaption><em>Table 1: Sample of records from example dataset. Null values indicate missing data.</em></figcaption></figure><p>To establish a baseline to compare the identity resolution algorithm performance, we used deterministic matching, which is typically done in naive identity resolution systems. This involved using the id_number as a unique identifier, something we have commonly seen in the travel industry. This method yields and accuracy of 49%, with a precision of 100%, recall of 49% and f1 score of 66%.</p><p>Results are also shown in experiment 1 from table 2. The results can be surmised by: deterministic matching won’t incorrectly match any records that don’t belong to the same person (precision) but is poor at finding actual records corresponding to people (recall). Using this method would result in a poor single customer view as it only captures half of the correct matches, meaning customer profiles would be sparsely populated and resulting analysis or modelling on them would be suboptimal.</p><p>Next, we passed this data through the Blueprint: Identity Resolution algorithm on AWS and used name blocking. As previously discussed in part 2 of the blog, blocking is a method used to reduce the computational load of the identity resolution process that arises due to the non-linear scaling of the algorithm. While blocking isn’t really required in this case due to low input data volumes, we want to simulate how the algorithm would be used in reality.</p><p>The exact block used was the first 2 letters of surname. The default values for fuzzy matching thresholds from Blueprint: Identity Resolution were used as well as the built-in matching rules:</p><ol><li>If id numbers fuzzy match, records belong to same person;</li><li>If given name, surname and date of birth fuzzy match, records belong to the same person;</li><li>If given name and date of birth fuzzy match, records belong to the same person.</li></ol><p>This resulted in an accuracy of 70%, with precision of 100% (note: this is a rounded value, meaning actual value could be 99.5% or greater), recall of 70% and f1 score of 82% (experiment 2 in table 2).</p><p>Another experiment (experiment 3 in table 2) used year of birth as the blocking key, keeping all parameters the same as experiment 2 and this had an accuracy of 76%, with precision of 99%, recall of 77% and f1 score of 86%. From these results, it’s apparent that the identity resolution algorithm shows significant improvement over the deterministic method, meaning customer profiles have been enriched, leading to positive effects on downstream analysis etc.</p><p>We note that we used all default values and pre-inbuilt rules for these experiments. Better results could be achieved by tuning fuzzy matching thresholds and having additional rules, more suited to this specific dataset but for convenience, we used the out of the box solution.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SNHgs9srNwKe3GkeqZYawg.png" /><figcaption>Table 2. <em>Results of experiments on simulated data using a variety of techniques.</em></figcaption></figure><p>To test how blocking effects performance metrics, an experiment using no blocking, meaning all records were compared to each other was conducted. We wouldn’t normally recommend this, unless the data is small or you have considerable computing resources. But since this is a small dataset, it wasn’t much of a problem. The exact same parameters from previous experiments were used and the algorithm had an accuracy of 81%, with precision of 98%, recall of 82% and f1 score of 89% (experiment 3 in table 2). This shows a reasonable increase on the blocking methods, whether it should be done in reality, very much depends on the conditions previously mentioned.</p><p>The results of the no blocking experiment got us thinking on how to improve the algorithm going forward. We decided to combine the results of exact matching, name and year of birth blocking experiments together (experiment 7 in table 2) to see how much lift is gained and how it compares to not blocking at all. The accuracy was 83%, with precision of 99%, recall of 83% and f1 score of 91%. This result shows some promise as a possible avenue for future development but when dealing with anything related to identity resolution, algorithm performance always has to be balanced with resources and acceptable run times. For instance, using 6 or 7 blocks would be a very heavy process and you would eventually get diminished returns, so an acceptable balance must be found. Other, similar experiments were run and the results are shown in experiments 5 and 6 in table 2.</p><p><em>Matching rules versus machine learning</em></p><p>The final experiment of interest was comparing how well matching rule techniques compare to a machine learning approach. As previously explained, there are pros and cons associated with using machine learning to find matches. The main one being the lack of labeled data, but in our case, since the data is synthetic, the data comes with labels. The modelling dataset consisted of 12.5M records (all possible combinations of the 5,000 input records) since no blocking was used. The positive match rate of the data was about 0.05%, making this a severe class imbalance problem. The data was split up into 3 sets, a training set (50%), a validation (20%) and test set (30%). The validation set was used for finding the optimal threshold value for deciding if 2 records belong to the same entity and the test data for model evaluation.</p><p>Since there was a severe class imbalance, the training set was down-sampled so that the positive class was about 5% of the data, rather than 0.05%. We used the logistic regression algorithm from the <a href="https://scikit-learn.org/stable/">scikit-learn</a> library with default settings (no hyper-parameter tuning). Logistic regression is a more than suitable modelling technique for this problem since it should be linear in nature, i.e. the more fields that are more similar have a higher probability of belonging to the same entity than not. As a result, we don’t expect any non-linear behaviour that this model couldn’t capture. It is also less data hungry and the results are more explainable that other methods such as boosting and neural networks.</p><p>The threshold for a positive match was chosen based on precision-recall curves on the validation set and the value of 84% was found to be optimal, this means the model had to be a minimum of 84% confident that two records belong to the same entity. Typically, in machine learning classifiers a default value of 50% is used, but here we want to be very sure that records match to minimise records belonging to separate entities to be linked together.</p><p>The results from the test set show the model was 97% accurate, with precision of 99%, recall of 97% and f1 score of 98% (experiment 8 in table 2). This clearly outperforms the rules based methods used previously but in some ways, it’s an unfair comparison. The rules used weren’t specific to this data like the model was, rather they were general rules that have worked for us in previous client implementations but weren’t specific to this dataset, whereas the model could capture more complicated and specific trends from the data. It would be possible to make specific rules for this dataset that would lead to increased performance. Nonetheless, it shows that if you have labelled data, a classifier might be the best option as it would probably outperform any user made rules. Also, it requires less decisions to be made by users, such as deciding fuzzy matching thresholds for each field that was compared and there is no need to make rules, as these raw numbers are inputs into the model.</p><p><strong>Conclusion</strong></p><p>The purpose of this three part article is to share with the data engineering and data science community in the travel industry how we perform identity resolution at OpenJaw to create a single-customer-view.</p><p>In <a href="https://medium.com/the-openjaw-data-science-blog/identity-resolution-for-travel-retailing-part-1-b16ff6d1edd6">part 1</a> of the article, we described why identity resolution is needed as well as some of the technical challenges associated with it, such as the nonlinear properties of the algorithm brought about by the large search space associated with finding matches. In <a href="https://medium.com/the-openjaw-data-science-blog/identity-resolution-for-travel-retailing-part-2-3b5752cd9930">part 2</a>, we described how we use data cleansing and normalisation, ‘blocking’ and ‘salting’ to prepare the data for fuzzy matching.</p><p>And finally, in this part of article, we continued on with our description of how the algorithm works by discussing how we use empirical matching rules and machine learning to match records and graph theory to link records that belong to the same entity. We also described the results of experiments that were conducted on publicly available data that compare our identity resolution algorithm performance to traditional, deterministic matching.</p><blockquote>These results showed that probabilistic matching greatly outperforms deterministic methods.</blockquote><p>Potential future improvements to the algorithm were also explored, such as: combining deterministic matching and probabilistic matching, creating a hybrid solution as well as how machine learning could outperform matching rules, once a labeled dataset is available.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cfc0ddfd2e87" width="1" height="1" alt=""><hr><p><a href="https://medium.com/the-openjaw-data-science-blog/identity-resolution-for-travel-retailing-part-3-cfc0ddfd2e87">Identity resolution for travel retailing (part 3)</a> was originally published in <a href="https://medium.com/the-openjaw-data-science-blog">The OpenJaw Data Science Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[In travel retailing, some customers are more equal than others (part 3)]]></title>
            <link>https://medium.com/the-openjaw-data-science-blog/in-travel-retailing-some-customers-are-more-equal-than-others-part-3-d2235c5fb1bc?source=rss-7bd43b8b9097------2</link>
            <guid isPermaLink="false">https://medium.com/p/d2235c5fb1bc</guid>
            <category><![CDATA[travel-retail-market]]></category>
            <category><![CDATA[clv]]></category>
            <category><![CDATA[customer-analytics]]></category>
            <category><![CDATA[marketing]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[The OpenJaw Data Science Team]]></dc:creator>
            <pubDate>Fri, 13 Dec 2019 09:55:43 GMT</pubDate>
            <atom:updated>2020-02-28T13:56:52.079Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*i7sXzGUWkSm7Uoa84pxQvA.jpeg" /></figure><p>In this article, we share how we predict Customer Lifetime Value (CLV) at OpenJaw and explain why this metric is so important for travel retailing. This is the third and final part of a three part article on this topic.</p><p><em>By </em><a href="https://www.linkedin.com/in/john-carney-40328920/"><em>John Carney</em></a><em>, </em><a href="https://www.linkedin.com/in/charles-wang-174970144/"><em>Yuxiao Wang</em></a><em>, </em><a href="https://www.linkedin.com/in/aurenferguson/"><em>Auren Ferguson</em></a><em>, </em><a href="https://www.linkedin.com/in/dr-beibei-flynn-cta-2681129/"><em>Beibei Flynn</em></a><em> and </em><a href="https://www.linkedin.com/in/gavin-kiernan-365a359/"><em>Gavin Kiernan</em></a><em>.</em></p><p><strong>Introduction</strong></p><p>In <a href="https://medium.com/the-openjaw-data-science-blog/in-travel-retailing-some-customers-are-more-equal-than-others-part-2-65276ebf2dbb">part 2</a> of this article, we described in detail how the Pareto/NBD model can be used for CLV prediction in travel retailing. In this, the third and final part of the article, we describe how Random Forest machine learning can also be used successfully for this purpose. We also compare the predictive performance of both methods and provide some guidance on which method to use in different circumstances.</p><p><strong>Using Random Forest regression to predict CLV</strong></p><p>As described in detail in <a href="https://medium.com/the-openjaw-data-science-blog/in-travel-retailing-some-customers-are-more-equal-than-others-part-2-65276ebf2dbb">part 2</a> of this article, the first step in predicting CLV is to predict what we call Customer Lifetime Spend (<em>CLS</em>). We define <em>CLS</em> as the product of Predicted Average Purchase Value (<em>PAPV</em>) and Predicted Number of Purchases (<em>PNP</em>):</p><p><em>CLS = PAPV </em>x<em> PNP</em></p><p>Once margin for a specific customer segment and product is available, it is applied to calculate <em>CLV:</em></p><p><em>CLV</em> = <em>CLS</em> x <em>Margin</em></p><p>Our approach in this article is to build two separate Random Forest regression models to estimate <em>PAPV </em>and <em>PNP. </em>An alternative approach is to build a single Random Forest model that predicts <em>CLS </em>directly. However, if we did this it would be more difficult to uncover where the performance gain or loss is in the Random Forest models versus the traditional models, so we decided for the purposes of the experiments in this article to model <em>PAPV </em>and <em>PNP</em> separately.</p><p><em>Training set</em></p><p>The training set that we use for the experiments described below is constructed from 270,000 flight transactions or Passenger Name Records (PNR’s). We process these transactions with our Identity Resolution algorithm, described in a <a href="https://medium.com/the-openjaw-data-science-blog/identity-resolution-for-travel-retailing-part-1-b16ff6d1edd6">separate article</a>. This creates a single customer view by assigning each transaction a unique customer identifier. Transactions are then rolled up to create features for each customer.</p><p><em>Training process</em></p><p>To train the Random Forest model, we start by picking a prediction window. This is the length of time in the future we want the model to predict into. For the experiments summarised below, we chose T = 1 year.</p><p>As illustrated in Figure 1 below, the full data set spanned 3 years. To construct a training set for model training, we take data from the most recent period (shown in red), calculate each customer’s number of purchases and average spend per booking as target values. We then join these to the corresponding input features (shown in blue). For customers that didn’t purchase in the ‘target’ period, their target values are set to zero.</p><p>Note that, for the <em>PAPV</em> model, only customers that actually purchased in the ‘Target’ period can be used to train the model, because <em>PAPV </em>is unobserved for people without purchases.</p><p>Once both models are trained (<em>PAPV </em>and <em>PNP</em>), we can apply the model on the most recent 2 years of data (shown in green), to estimate <em>PCLS </em>for the future (shown in yellow).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kdKpj0Nc-VdkPgHx085mWA.png" /><figcaption>Figure 1: Model training and predicting time-scales.</figcaption></figure><p><em>Feature selection</em></p><p>Our training set contained a relatively large number of features (62), so we utilised a feature selection process to identify the most predictive features and separate them from features that are adding little predictive value, or are just noise.</p><p>The first step in this process is to drop a feature if more than 20% of the field values are null, or the variance of the field is less than 10% of the mean value. The second step is to apply Random Forest ‘feature importance’ on all of the remaining variables. This method calculates how much each feature decreases the target variance in a tree, weighted by the probability of a customer / training vector reaching that tree. It then accumulates this over all trees in the Random Forest. This is known as the Mean Decrease in Impurity (MDI) method (2).</p><p>Tables 1 and 2 below show the important features for the <em>PNP</em> and <em>PAPV</em> models. The importance scores are normalised so that they sum to 1 and we drop features with importance scores less than or equal to 0.001.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TG69bZpGjPOAY20DAsSvnQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zLRCm3EQdjttu-F9ciHoFw.png" /></figure><p>As expected, the Recency, Frequency and Monetary (RFM) related features are ranked highly, as it is highly probable that the Random Forest model captures the same underlying process of customer purchasing behaviour that the traditional RFM based models like Pareto/NBD capture. Other features, such as ‘Number of bookings with 1 adult passenger only’ or ‘Number of bookings with same day return’ may look surprising at first glance, but we know at OpenJaw that these features correspond to business travel and can provide additional information regarding the Frequency of a customer.</p><p>Further analysis is required to form definitive conclusions on this, but it does seem that the Random Forest model is doing something very similar to traditional RFM based models for predicting CLV, but enriching itself with additional features that contain more information on RFM (versus improving performance with completely new, exogenous features).</p><p><strong>Experiments</strong></p><p>To benchmark the prediction performance of our Random Forest method, we compared it to two established RFM based methods; BG/NBD (4) and the Pareto/NBD model described in part 2 of this article. We also wanted to measure the value of including the additional features listed in tables 1 and 2 above, so we compared the prediction performance of four models in total:</p><ol><li>Random Forest regression model with the features listed in tables 1 and 2 above;</li><li>Random Forest regression model with just three features; Recency, Frequency and Spend;</li><li>BG/NBD model. This works with just Recency, Frequency and Spend;</li><li>Pareto/NBD model. This works with just Recency, Frequency and Spend.</li></ol><p>The prediction performance of each model on the test set, expressed in (normalised) Root Mean Squared Error (RMSE), is summarised in Table 3 below. The best performing models are highlighted in green text.</p><p>The Pareto/NBD model wins overall on <em>CLS</em> prediction as it outperforms other models on <em>PNP</em> prediction significantly, while the Random Forest model with additional features appears to be better at <em>PAPV</em> predicting.</p><p>Note that both the BG/NBD and Pareto/NBD have the same RMSE for <em>PAPV </em>prediction. This is because they use the same process for prediction (based on the Gamma-Gamma distribution assumption).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/939/1*oicmP9py0ZI3CjutGYADbQ.png" /></figure><p>So which model is best? The <em>CLS </em>prediction performance would suggest that the Pareto/NBD model is the best performer overall, followed by BG/NBD and then the Random Forest models. However, it is worthwhile taking a deeper look into how this performance is distributed for different types of customers. Figure 2 below illustrates how <em>CLS</em> prediction performance varies across the test set for customers with increasing frequency of purchase.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/939/1*7rNC62XP1FFs_eywRBQiBg.png" /><figcaption>Figure 2. Prediction error (RMSE%) for customers with increasing Frequency in the test set.</figcaption></figure><p>As we expected, the prediction performance of the Random Forest models improve notably when Frequency is larger than 2. This is because machine learning methods like Random Forest are generally non-parametric i.e. they have no distributional (parametric) assumptions to fall back on when the number of observations are small for a particular type of customer. In contrast, the traditional RFM models are parametric, so can continue to perform reasonably well, even when the number of observations are small.</p><p><strong>Conclusion</strong></p><p>The purpose of this three part article is to share with the data science and marketing community in the travel industry how we predict CLV at OpenJaw and also to explain why this metric is so important for travel retailing.</p><p>The key message of <a href="https://medium.com/the-openjaw-data-science-blog/in-travel-retailing-some-customers-are-more-equal-than-others-part-1-202eb78d869f">part 1</a> of the article is that CLV provides a much more precise measure of customer value relative to the informal methods used today in travel e.g. points accumulated in a loyalty program or the fare class for a particular trip. Another key message it that the field of CLV is somewhat of a ‘minefield’, with many variants and a myriad of vague publications online that confuse one approach over another. So caution is advised when embarking on a journey towards CLV adoption.</p><p>In <a href="https://medium.com/the-openjaw-data-science-blog/in-travel-retailing-some-customers-are-more-equal-than-others-part-2-65276ebf2dbb">part 2</a> of the article we focused on the most established method used to predict CLV in a non-contractual setting such as travel; the Pareto/NBD method. We explained the mechanics of how it works and also provided some examples with real travel data to illustrate how the process of customer purchasing behaviour is captured by the model.</p><p>And finally, in part 3 of the article, we described how the popular machine learning method Random Forest can also be used to predict CLV. To evaluate its performance, we compared it to two traditional methods, Pareto/NBD and BG/NBD. The prediction performance of each model was similar, but for higher frequency customers, the Random Forest model with additional features beyond RFM, performed consistently better. This result provides some guidance on which model to use in different circumstances and also illustrates the potential of machine learning for CLV prediction in travel.</p><p><strong>References</strong></p><p>(1) Chamberlain, B. P., Cardoso, A., Liu, C. H., Pagliari, R., &amp; Deisenroth, M. P. (2017, August). Customer lifetime value prediction using embeddings. In <em>Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining</em> (pp. 1753–1762). ACM.</p><p>(2) Friedman, J., Hastie, T., &amp; Tibshirani, R. (2001). <em>The elements of statistical learning</em> (Vol. 1, №10). New York: Springer series in statistics.</p><p>(3) Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Vanderplas, J. (2011). Scikit-learn: Machine learning in Python. <em>Journal of machine learning research</em>, <em>12</em>(Oct), 2825–2830.</p><p>(4) Fader, P., Hardie, B., Lee, K. (2005). Counting your customers the easy way: An alternative to the Pareto / NBD model. <em>Marketing Science, </em>Vol. 24, №2, Spring 2005, pp. 275–284.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d2235c5fb1bc" width="1" height="1" alt=""><hr><p><a href="https://medium.com/the-openjaw-data-science-blog/in-travel-retailing-some-customers-are-more-equal-than-others-part-3-d2235c5fb1bc">In travel retailing, some customers are more equal than others (part 3)</a> was originally published in <a href="https://medium.com/the-openjaw-data-science-blog">The OpenJaw Data Science Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Welcome to the OpenJaw Data Science Blog!]]></title>
            <link>https://medium.com/the-openjaw-data-science-blog/welcome-to-the-openjaw-data-science-blog-5a14181ca29?source=rss-7bd43b8b9097------2</link>
            <guid isPermaLink="false">https://medium.com/p/5a14181ca29</guid>
            <category><![CDATA[big-data-analytics]]></category>
            <category><![CDATA[travel]]></category>
            <category><![CDATA[customer-data-platform]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[The OpenJaw Data Science Team]]></dc:creator>
            <pubDate>Mon, 04 Nov 2019 09:36:09 GMT</pubDate>
            <atom:updated>2019-11-04T09:36:09.019Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TUFzAUVcq-Y6FGC3q6qC5w.png" /></figure><p>Insights and opinions for data scientists and engineers in the travel industry.</p><p><em>By </em><a href="https://www.linkedin.com/in/john-carney-40328920/"><em>John Carney</em></a><em>, Chief Data Scientist, OpenJaw Technologies</em></p><p><a href="http://www.openjawtech.com/">OpenJaw</a> started its data science journey on 1st May 2017. I remember it well because it was my first day at the company! Since then, we have hired a world class team of data scientists and data engineers, built a petabyte scale Big Data platform for customer analytics and personalisation in travel and deployed this platform into production for some of the largest and most innovative travel brands in the world. We are now, in October 2019, generating insights for personalised travel retailing for over 40 million travellers globally; quite an achievement from a standing start just over 2 years ago.</p><p>The purpose of this blog is to openly share the lessons learned and technical insights we have accumulated over the past 2 years on our journey. We will continue to do this as we develop our platform into the future. We want to give something back to the data science community in travel to accelerate progress in the field and also to engage with the field in an open, pro-active and collaborative way. If you are working as a data scientist in the travel industry, or in the field of personalisation in any industry, it is highly probable that you are currently trying to solve the same or very similar problems as we are here at OpenJaw.</p><p>In our first series of articles, we will share with our readers the methods we use to accurately predict customer lifetime value, perform identity resolution, predict customer purchase propensity and segment customers with machine learning — all ‘pillars’ of personalisation for any online retailer or advertiser today. In the area of data engineering, we will focus on the key topics of data pipeline management, ETL and data warehousing, spanning technologies such as Apache Spark, Hadoop and Airflow. We will also touch on the challenges of data visualisation at scale and how to optimise Big Data deployments in the Cloud.</p><p>More than anything though, this blog is a testament to the amazing technical creativity, determination and hard-work of the data team here at OpenJaw. We have some emerging stars of data science and data engineering here and I think you will enjoy listening to what they have to say and taking the time to understand how they think and solve the very challenging data science and Big Data engineering problems they face every day.</p><p>Most of our articles are written in consumable parts of about 1,000 words each, but we will endeavour to cross-reference them so that readers can see how everything fits into coherent vision for data science in travel.</p><p>I do hope you enjoy our articles. Feel free to reach out to us with feedback and questions. And don’t forget to clap! Happy reading!</p><p><a href="https://www.linkedin.com/in/john-carney-40328920/"><em>John Carney</em></a><em>, Editor, The OpenJaw Data Science Blog</em></p><p>Data.Science@openjawtech.com</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/457/1*4l-jOH9JFCTgTYLWHKo3vw.jpeg" /><figcaption>Dr. John Carney, Chief Data Scientist, OpenJaw Technologies</figcaption></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5a14181ca29" width="1" height="1" alt=""><hr><p><a href="https://medium.com/the-openjaw-data-science-blog/welcome-to-the-openjaw-data-science-blog-5a14181ca29">Welcome to the OpenJaw Data Science Blog!</a> was originally published in <a href="https://medium.com/the-openjaw-data-science-blog">The OpenJaw Data Science Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Identity resolution for travel retailing (part 2)]]></title>
            <link>https://medium.com/the-openjaw-data-science-blog/identity-resolution-for-travel-retailing-part-2-3b5752cd9930?source=rss-7bd43b8b9097------2</link>
            <guid isPermaLink="false">https://medium.com/p/3b5752cd9930</guid>
            <category><![CDATA[travel]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[retail-marketing]]></category>
            <category><![CDATA[identity-resolution]]></category>
            <category><![CDATA[graph-theory]]></category>
            <dc:creator><![CDATA[The OpenJaw Data Science Team]]></dc:creator>
            <pubDate>Sun, 03 Nov 2019 14:21:01 GMT</pubDate>
            <atom:updated>2019-12-03T17:19:17.832Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tvatfoH6jBWTcO-Rs2opUg.jpeg" /><figcaption>An identity graph, similar to those created by OpenJaw’s data science team for identity resolution.</figcaption></figure><p>In this article, we share how we create a single-customer-view at OpenJaw using a new identity resolution algorithm we have developed which is optimised for the travel domain. This is the second part of a three-part article on this topic.</p><p><em>By </em><a href="https://www.linkedin.com/in/aurenferguson/"><em>Auren Ferguson</em></a><em>, </em><a href="https://www.linkedin.com/in/john-carney-40328920/"><em>John Carney</em></a><em>, </em><a href="https://www.linkedin.com/in/vignesh-mohan94/"><em>Vignesh Mohan</em></a><em>, </em><a href="https://www.linkedin.com/in/gavin-kiernan-365a359/"><em>Gavin Kiernan</em></a><em>, </em><a href="https://www.linkedin.com/in/charles-wang-174970144/"><em>Yuxiao Wang</em></a><em> and </em><a href="https://www.linkedin.com/in/dr-beibei-flynn-cta-2681129/"><em>Beibei Flynn</em></a><em>.</em></p><h4><strong>Introduction</strong></h4><p>In <a href="https://medium.com/the-openjaw-data-science-blog/identity-resolution-for-travel-retailing-part-1-b16ff6d1edd6">part 1</a> of this article, we described the key challenges of identity resolution in the travel domain. In this part, we will describe some of the details of the algorithm we have developed at OpenJaw to solve this problem.</p><h4><strong>Data cleansing</strong></h4><p>It’s commonly mentioned that about 80% of a data scientist’s time is tied up with data cleansing. Identity resolution is no different in this regard. Cleansing and standardising the data is a fundamental and, unfortunately, very time consuming part of the process.</p><p>The first step is to get the data from all sources into a common schema. At OpenJaw, we have devised a flexible template schema that standardises field names and data types, for example:</p><ul><li>given_name: string</li><li>surname: string</li><li>date_of_birth: integer</li><li>booking_reference: string</li></ul><p>We then combine the data from all sources into a single dataset and proceed to cleanse it. The cleansing process can vary massively from dataset to dataset but generally would involve steps such as:</p><ul><li>Convert all string fields to lower case</li><li>Remove accents from names</li><li>Remove/add phone number country codes</li><li>Remove non-numerics from phone numbers</li><li>Extract information such as gender and date of birth from identification documents</li></ul><p>While this stage is often tedious and labour intensive, the goal is to get the data in good shape, ready to pass through the rest of the process seamlessly.</p><h4><strong>Reducing computational overhead with ‘blocking’</strong></h4><p>As mentioned in part 1 of this article, the computational expense of identity resolution is extremely high. The next steps are to reduce the number of potential matches by data reduction/de-duplication and a method called ‘blocking’.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kAMU6bI24xTM6OGdVBs0uQ.png" /><figcaption>Data from three sources are standardised, combined into one format and cleansed.</figcaption></figure><p>For data reduction, we assign each vector of personal identifier information a unique reference number. Depending on the degree of duplication, this can greatly cut down the number of records that need to be compared. For example, if a frequent customer (&gt; 100 bookings) books using the same details each time, we can reduce that number down to 1 record, or as we call it; a persona version. A person/entity could have 10’s or 100’s of these personas and the job of identity resolution is to link them together.</p><p>Blocking is a method used to reduce the overall number of comparisons between records during the identity resolution process. It restricts the number of comparisons by creating subsets of the data by a certain criteria and only performs full matching between records within the subset. Blocking methods have been studied extensively in academia (see e.g. (1)), with methods ranging from very simple blocks, such as postcode to more complicated methods like phonetics of surname (similarly sounding surnames). Which blocking techniques to use will depend on the given data and there is no ‘correct’ block that will capture all matches in the same group.</p><p>To frame it from a machine learning perspective, choosing a block is a precision-recall problem. We want to balance getting as many potential matches into the same block but also dramatically reduce the overall number of comparisons, reducing computational load. A simple block we have used before is the first letter of the first name and the first two letters of surname. The rational being that when filling out online forms, people are far less likely to make mistakes/typos at the beginning of their names.</p><p>Blocking will always exclude matches, using the previous example, what if someone gets married and chooses to change their surname? The name block mentioned won’t work. As a result, it’s common to use multiple blocks to reduce this error, e.g. use name block and the year from date of birth, as shown in the figure below. There are two entities in this example, William (Will/William) and Mary. Using the name block, we capture all of William’s records (block ‘ws’) but not Mary’s (block ‘mt’ and ‘mp’), since she recently got married and changed her name. The second block captures her records since the year of birth is the same. Since row 3 has no date of birth the relationship between it and row 1 and 2 are missed. The different candidate matches from different blocks are combined and deduplicated to create the final candidate matches. These are the records that will be ‘fuzzy matched’ later. The number of comparisons has reduced from 10, in the case of no blocking, to 4 comparisons using blocking, yet captures all the information required to resolve identities.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/869/1*auHLsnmBydqQjKeypIBkIA.png" /><figcaption><em>Blocking process using two methods, the first method uses the first letter of given name and surname and the second block is the year of birth.</em></figcaption></figure><p>The next stage of the process is the most technically and computationally challenging part as it involves the creation of the candidate matches on which fuzzy matching will be done. This involves a very large self-join on the blocking field(s) of the input data that balloons the data volume dramatically. In a typical implementation at OpenJaw, we may have as many as 100M records to resolve, which corresponds to approx 250 billion candidate matches. This is compounded by the fact that many of the datasets we work with at OpenJaw are highly skewed, with many blocks having over 100,000 records, but most having only 10–100’s of records.</p><p>To handle this complexity and make this a tractable problem computationally, we developed the following method:</p><p>1) We ‘salt’ the left version of the dataset so it’s evenly distributed between nodes/executors (2). This involves adding a degree of randomness to the join key. In this case a number between 1 and 1000 is appended the blocking field e.g the block is birth year (1980), salted key becomes 1980123. We can then repartition the dataset on this key to make it evenly distributed.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*T3GwdlIxE_Qbl4B3dvZICA.jpeg" /><figcaption>‘Salting’ identity data is an important part of identity resolution at OpenJaw.</figcaption></figure><p>2) We break the right dataset (i.e. a non-salted copy of the left dataset) into 2 parts, one with skewed blocks (this is an arbitrarily large minimum block size depending on data volume) and one without skewed blocks. The skewed dataset is broken up into a number of chunks, with each chunk having all records of a block and approximately have the same number of comparisons to calculate.</p><p>3) We perform iterative broadcast joins of each chunk to the left dataset. This removes the skew and each executor takes around the same time to complete.</p><p>4) We perform a regular sort merge join on the un-skewed data and combine them back into one dataset that is ready for fuzzy matching.</p><p>Using this method, as well as creating persona profiles as mentioned above, reduced the run time from over a weekend to just under an hour!</p><h4><strong>Measuring string similarities with fuzzy matching</strong></h4><p>At this stage, the data is ready for fuzzy matching algorithms to measure string similarities between fields of candidate matches. There is a whole host of string similarity metrics. For identity resolution we are primarily concerned with edit distance metrics, that is to say, metrics that compute the number of operations needed to transform one string to another.</p><p>In our solution, we use a combination of exact matching for fields such as passport and national ID numbers, ‘Jaro-Winkler’ (3) distance for names (since it places more importance to differences at the beginning of strings where differences are more likely to be real rather than typos, e.g. the assumption that name typos will be towards the end of a name) and ‘Levenshtein’ (4) distance for phone numbers, emails, loyalty numbers, etc.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/632/1*wc3vnJpISumnCCBND1oHMQ.png" /><figcaption>Comparison of a name and phone number using different similarity metrics.</figcaption></figure><p>The aim of these algorithms is to quantify how close strings are to each other; if the similarity score is 0, they are considered completely different, whereas a score of 1 is equivalent to an exact match. The table below provides some examples of typical outputs from these matching algorithms. As is evident in the results, Jaro-Winkler is generally preferred with names since it weights the beginnings of strings more, whereas Levenshtein distance shows good performance on strings of numbers.</p><p>To recap, we have measured the similarities between fields for all blocks in the data, where its typical to fuzzy match between 5–10 fields for each candidate match. The next step is to decide if each candidate match does or doesn’t belong to the same entity.</p><h4><strong>In part 3 of the article…</strong></h4><p>In the next part of this article, we will describe how we use a combination of empirical matching rules, machine learning and graph theory to complete the identity resolution process and create a single-customer-view.</p><h4><strong>References</strong></h4><p>(1) Steven Euijong Whang, David Menestrina, Georgia Koutrika, Martin Theobald, and Hector Garcia-Molina. 2009. Entity resolution with iterative blocking. In <em>Proceedings of the 2009 ACM SIGMOD International Conference on Management of data (SIGMOD ‘09)</em>, Carsten Binnig and Benoit Dageville (Eds.). ACM, New York, NY, USA, 219–232. DOI: <a href="https://doi.org/10.1145/1559845.1559870">https://doi.org/10.1145/1559845.1559870</a></p><p>(2) (a) <a href="https://bigdatacraziness.wordpress.com/2018/01/05/oh-my-god-is-my-data-skewed/">https://bigdatacraziness.wordpress.com/2018/01/05/oh-my-god-is-my-data-skewed/</a>. (b) <a href="https://dzone.com/articles/why-your-spark-apps-are-slow-or-failing-part-ii-da">https://dzone.com/articles/why-your-spark-apps-are-slow-or-failing-part-ii-da</a>. (c) <a href="https://medium.com/appsflyer/salting-your-spark-to-scale-e6f1c87dd18">https://medium.com/appsflyer/salting-your-spark-to-scale-e6f1c87dd18</a></p><p>(3) <a href="https://en.wikipedia.org/wiki/Jaro%E2%80%93Winkler_distance">https://en.wikipedia.org/wiki/Jaro%E2%80%93Winkler_distance</a></p><p>(4) h<a href="https://en.wikipedia.org/wiki/Levenshtein_distance">ttps://en.wikipedia.org/wiki/Levenshtein_distance</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3b5752cd9930" width="1" height="1" alt=""><hr><p><a href="https://medium.com/the-openjaw-data-science-blog/identity-resolution-for-travel-retailing-part-2-3b5752cd9930">Identity resolution for travel retailing (part 2)</a> was originally published in <a href="https://medium.com/the-openjaw-data-science-blog">The OpenJaw Data Science Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[In travel retailing, some customers are more equal than others (part 2)]]></title>
            <link>https://medium.com/the-openjaw-data-science-blog/in-travel-retailing-some-customers-are-more-equal-than-others-part-2-65276ebf2dbb?source=rss-7bd43b8b9097------2</link>
            <guid isPermaLink="false">https://medium.com/p/65276ebf2dbb</guid>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[The OpenJaw Data Science Team]]></dc:creator>
            <pubDate>Sat, 02 Nov 2019 14:36:02 GMT</pubDate>
            <atom:updated>2019-11-05T15:33:43.104Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*51HoMmEpPbNjcjQnssUdpg.jpeg" /></figure><p>In this article, we share how we predict Customer Lifetime Value (CLV) at OpenJaw and explain why this metric is so important for travel retailing. This is the second part of a three part article on this topic.</p><p><em>By </em><a href="https://www.linkedin.com/in/john-carney-40328920/"><em>John Carney</em></a><em>, </em><a href="https://www.linkedin.com/in/charles-wang-174970144/"><em>Yuxiao Wang</em></a><em>, </em><a href="https://www.linkedin.com/in/aurenferguson/"><em>Auren Ferguson</em></a><em>, </em><a href="https://www.linkedin.com/in/dr-beibei-flynn-cta-2681129/"><em>Beibei Flynn</em></a><em> and </em><a href="https://www.linkedin.com/in/gavin-kiernan-365a359/"><em>Gavin Kiernan</em></a><em>.</em></p><h4><strong>Introduction</strong></h4><p>In the first part of this article, we provided an introduction to the underlying concept of CLV and explained the practical challenges that arise when applying CLV to the travel domain. In this part of the article we describe in detail one of the two methods we use at OpenJaw for CLV prediction; the ‘Pareto/Negative Binomial Distribution (NBD)’ model.</p><h4>Suitable methods for predicting CLV in travel</h4><p>There are two primary quantitative methods suitable for predicting CLV in a non-contractual setting such as travel today.</p><p>The first method uses non-linear stochastic models of buyer behaviour to link the popular RFM (Recency, Frequency, Monetary) paradigm with CLV. There are two key variants of this approach — the Beta-Geometric/Negative Binomial Distribution (BG/NBD) model first proposed in (2) and the Pareto/NBD model first proposed in (1). Generally speaking, the Pareto/NBD model is more accurate, but has a higher computational overhead than BG/NBD and is more complex to implement.</p><p>The second method uses machine learning methods such as <a href="https://en.wikipedia.org/wiki/Random_forest">Random Forest</a> to predict CLV. An early example of such an approach applied in a non-contractual setting such as travel is described in (3). This approach normally aims to create models that combine the usual RFM features with exogenous features such as customer demographic information to improve predictive performance.</p><p>Neither of these methods can claim to be the ‘best’ method for CLV prediction in travel in all circumstances; this really depends on how much data you have, its quality, as well as how much computing power you have available to parametrise / train the models.</p><p>Generally, at OpenJaw, our approach is to use Pareto/NBD when the history of RFM data is limited — parametric methods such as Pareto/NBD are less likely to overfit their training set, so tend to generalise better in this scenario. On the other hand, if we have access to a large history of RFM data as well as exogenous information such as demographic profiles for every customer, then we normally use the machine learning method with Random Forest.</p><p>Before deploying in a real-world scenario, we always confirm which method is best for a particular customer data-set using <a href="https://en.wikipedia.org/wiki/Cross-validation_(statistics)">cross-validation</a> or a similar model validation technique. We describe the details of this experimental process in part 3 of this article.</p><h4><strong>The Pareto/NBD model: How does it work?</strong></h4><p>Commonly known as a ‘Buy Till You Die’ (BTYD) model, at the heart of this method is a framework that models the flow of customer transactions over time in a non-contractual setting:</p><blockquote>BTYD models jointly model two processes; a <strong><em>repeat</em></strong><em> purchase process</em> that explains how frequently customers make purchases while they are still ‘alive’ and a <em>dropout process</em> that models how likely it is that a customer is about to churn in any given period.</blockquote><p>The variant of Pareto/NBD that we use at OpenJaw is the version described in (2). This requires four pieces of information that are pertinent to customer purchasing behaviour:</p><ul><li>Recency: The duration of time in days between a customer’s first booking and most recent booking;</li><li>Frequency: The total number of repeat bookings the customer has made. This is equivalent to the total number of booking minus one;</li><li>Monetary: The total amount spent divided by the number of bookings made by a customer. Also called Average Purchase Value;</li><li>Time: The duration between a customer’s first booking and the end of the period of consideration.</li></ul><p>The model works by making assumptions regarding following:</p><ul><li>The number of purchases by any individual customer is Poisson distributed, with latent parameter 𝜆;</li><li>True customer lifetime is not observed but follows an Exponential distribution with latent parameter 𝜇;</li><li>Purchase value is Gamma distributed with scale parameter 𝜈;</li><li>The number of purchases and purchase value are mutually independent and customers are independent of one another.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*11abpYoZm_Q7fAUpM8hHHQ.png" /><figcaption>Statistical distributions assumed for customer purchasing behaviour in the Pareto/NBD Gamma-Gamma model for predicting CLV.</figcaption></figure><p>Building upon the above assumptions, the model makes the following ‘prior’ assumptions:</p><ul><li>True 𝜆 is unobserved and Gamma distributed;</li><li>True 𝜇 is unobserved and follows a Gamma distribution;</li><li>𝜆 and 𝜇 vary independently across customers;</li><li>The scale parameter of purchase value, 𝜈, is itself Gamma distributed.</li></ul><p>After working through the equations, which are detailed in (4), the following ‘posterior’ distributions apply:</p><ul><li>Customer lifetime follows a Pareto distribution;</li><li>The number of purchases by a customer follows a Negative Binomial Distribution (NBD);</li><li>Purchase value is ‘Gamma-Gamma’ distributed.</li></ul><p>This leads to the ‘Pareto/NBD Gamma-Gamma’ model for predicting CLV.</p><h4><strong>Using the Pareto/NBD Gamma-Gamma model to predict CLV</strong></h4><p>At OpenJaw, the first step in predicting CLV is to predict what we call Customer Lifetime Spend (CLS). Think of this as CLV, but without margin factored into the calculation.</p><p>The reason we do this is because, as described in part 1 of this article, in travel retailing, profit margin is highly variable at the customer level, even for similar products. To handle this complexity in a pragmatic way, our software enables end-users to apply margin to a CLS calculation after a customer segment and product has been selected.</p><p>We define the Predicted CLS (<em>PCLS</em>) as the product of Predicted Average Purchase Value (<em>PAPV</em>) and Predicted Number of Purchases (<em>PNP</em>):</p><p><em>PCLS = PAPV </em>x<em> PNP</em></p><p>Once margin for a specific customer segment and product is available, it is applied to calculate Predicted CLV (<em>PCLV):</em></p><p><em>PCLV</em> = <em>PAPV</em> x <em>PNP</em> x <em>Margin</em></p><h4><strong>Illustrative example</strong></h4><p>To illustrate how the Pareto/NBD Gamma-Gamma model works with real customer data, we will now describe how the model handles three typical purchase journeys for a low-cost airline and customer of OpenJaw.</p><p>Note that our goal here is to give readers an intuitive ‘feel’ for how our model works. We do this by showing how the ‘probability of being alive’ (P-alive) evolves over time for three typical customer purchase journeys; frequent customer, repeat customer and churned customer. Note that P-alive is essentially the probability a customer remains active, so is a key component of the <em>PNP</em> variable described above.</p><p>The underlying RFM data was extracted from multiple sources of Passenger Name Record (PNR) data. This represents the vast majority of transactional history (over 90%) for that airline across direct and indirect sales channels spanning 3 years.</p><p>Extracting the RFM data from the PNR data was non-trivial. This was because the personal data used to identify individual customers was inconsistently described in the PNR’s — this is a normal feature for this type of data. For example, some PNR’s had email addresses, others just had national ID’s, most had phone numbers but with prefixes that were inconsistently applied and so on.</p><p>To tackle this problem, OpenJaw’s identity resolution algorithm was used to create a single customer view which produces a unique identifier for every customer, together with a validated set of transactions. This was then used to accurately calculate Recency, Frequency and Monetary (spend) for every customer. Full details of this process and algorithm are published in another article in the OpenJaw Data Science Blog.</p><p><em>Scenario 1: Frequent customer</em></p><p>As illustrated in the graph below, a frequent customer maintains a high probability of repeat purchases as long as their (frequent) pattern of purchase is maintained. As soon as this pattern is disrupted, the P_alive probability decays rapidly because the customer is starting to fall outside a pattern of frequency as captured by the model.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MKLsGTXWMc7tmFZVbNfwBg.png" /><figcaption>The evolution of ‘probability of alive’ for a frequent airline customer.</figcaption></figure><p><em>Scenario 2: Repeat customer</em></p><p>In contrast to a frequent customer, a repeat customer that purchases infrequently will have a lower decay rate. This is illustrated below; the customer has only 2 purchases which are 8 months apart, so the decay rate for P_alive is lower. This is intuitive; you would expect to see that a consistently frequent customer will purchase again within a similar timeframe, while for other customers who rarely purchase, you won’t have the same expectation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sT_K7G4BoRsd2mFrAg8O0w.png" /><figcaption>The evolution of ‘probability of alive’ for a repeat airline customer.</figcaption></figure><p><em>Scenario 3: Churned customer</em></p><p>In the scenario illustrated below, the customer is frequent, but then stops purchasing. After about one year, the P_alive probability has reduced to less than 50%, so the customer has effectively churned. However, after 2 years the customer purchases again, so the model is reactivated and starts again.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0OPufagj7SaV0W1LMCQykg.png" /><figcaption>The evolution of ‘probability of alive’ for a churned airline customer.</figcaption></figure><h4><strong>In part 3 of this article…</strong></h4><p>In the next part of this article we will describe the Random Forest machine learning method we use to predict CLV and provide some empirical analysis that compares its predictive performance relative to the Pareto/NBD Gamma-Gamma model described above.</p><h4><strong>References</strong></h4><p>(1) Schmittlein, David C. and Peterson, Robert A. (1994). “Customer Base Analysis: An Industrial Purchase Process Application”, <em>Marketing Science, </em><strong>13 </strong>(Winter), 41–67.</p><p>(2) Fadar, Peter S.; Hardie, Bruce G.S.; Lee, Ka Lok (2005). “Counting Your Customers the Easy Way: An Alternative to the Pareto / NBD Model”, <em>Marketing Science, </em><strong>24</strong> (Spring), 275–284.</p><p>(3) Ali Vanderveld, Addhyan Pandey, Angela Han, and Rajesh Parekh (2016). “An Engagement-Based Customer Lifetime Value System for E-commerce”, <em>Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining.</em></p><p>(4) Peter S. Fader, Bruce G. S. Hardie, and Ka Lok Lee. 2005. “RFM and CLV: Using Iso-Value Curves for Customer Base Analysis”, <em>Journal of Marketing Research XLII</em>, November (2005), 415–430.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=65276ebf2dbb" width="1" height="1" alt=""><hr><p><a href="https://medium.com/the-openjaw-data-science-blog/in-travel-retailing-some-customers-are-more-equal-than-others-part-2-65276ebf2dbb">In travel retailing, some customers are more equal than others (part 2)</a> was originally published in <a href="https://medium.com/the-openjaw-data-science-blog">The OpenJaw Data Science Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Identity resolution for travel retailing (part 1)]]></title>
            <link>https://medium.com/the-openjaw-data-science-blog/identity-resolution-for-travel-retailing-part-1-b16ff6d1edd6?source=rss-7bd43b8b9097------2</link>
            <guid isPermaLink="false">https://medium.com/p/b16ff6d1edd6</guid>
            <category><![CDATA[identity-resolution]]></category>
            <category><![CDATA[hashing-and-salting]]></category>
            <category><![CDATA[big-data]]></category>
            <category><![CDATA[travel]]></category>
            <category><![CDATA[graph-theory]]></category>
            <dc:creator><![CDATA[The OpenJaw Data Science Team]]></dc:creator>
            <pubDate>Fri, 01 Nov 2019 09:59:55 GMT</pubDate>
            <atom:updated>2019-12-03T17:17:18.075Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OkRXAHlT42b6nNshB1HdTw.jpeg" /></figure><p>In this article, we share how we create a single-customer-view at OpenJaw using a new identity resolution algorithm we have developed which is optimised for the travel domain. This is the first part of a three-part article on this topic.</p><p><em>By </em><a href="https://www.linkedin.com/in/aurenferguson/"><em>Auren Ferguson</em></a><em>, </em><a href="https://www.linkedin.com/in/john-carney-40328920/"><em>John Carney</em></a><em>, </em><a href="https://www.linkedin.com/in/vignesh-mohan94/"><em>Vignesh Mohan</em></a><em>, </em><a href="https://www.linkedin.com/in/gavin-kiernan-365a359/"><em>Gavin Kiernan</em></a><em>, </em><a href="https://www.linkedin.com/in/charles-wang-174970144/"><em>Yuxiao Wang</em></a><em> and </em><a href="https://www.linkedin.com/in/dr-beibei-flynn-cta-2681129/"><em>Beibei Flynn</em></a><em>.</em></p><h4><strong>Introduction</strong></h4><p>In the travel sector, customer data is stored and managed across a multitude of different systems, spanning Passenger Servicing Systems (PSS), e-commerce systems, Customer Relationship Management (CRM) systems, Loyalty systems, digital marketing systems and, indirectly, via external systems such as social media platforms. At OpenJaw we can consume data from all of these sources, integrate it and use machine learning to generate insights on customer behaviour that can be used for personalisation.</p><p>The first step in this journey is to create a single-customer-view. The problem with combining customer data from a disparate set of data sources is that rarely is there a common, unique key that links all records that belongs to an entity. This is why identity resolution is required, as it provides a probabilistic / machine learning approach to linking entities from disparate data sources together. It is also common for identity resolution to be performed within one data source to link transactions within one system together.</p><p>A common example of where identity resolution is required at OpenJaw is that when customers book flights, hotels or vehicles on our e-commerce platform, they can use different combinations of personal information. For example, a person may use a passport as an ID when flying internationally but uses a driver’s licence or national ID when flying domestically. Usually, a passport or national ID is as close to a unique identifier a person can have, but in this instance, using conventional strict matching, these two flights will not be considered as being purchased by the same person. However, there could be similar information such as name, email and phone number(s) that could be used to match them with a high degree of probability using fuzzy matching, machine learning and graph algorithms.</p><p>Since our end goal is to deliver highly personalised customer experiences, it is imperative to be able to link customer records together in a reliable way. An analogy we like to use is that implementing identity resolution in personalisation is like building the foundations of a house; it is necessary to ensure that everything that is built after it performs as expected (e.g. predictive models of customer behaviour) and it is necessary to ensure customers are receiving relevant personalised offers that are aligned with marketing consents they have provided under data protection laws such as GDPR.</p><h4><strong>Key challenges in implementing identity resolution in travel</strong></h4><p>Conceptually, identity resolution is very simple to understand. But it is notoriously difficult to implement, especially in travel. This is due to a myriad of factors. For example, many of the airlines we work with at OpenJaw transport over 50 million passengers a year. When you collect this data over several years, it may resolve to over 100 million unique customer profiles. When you also consider the multitude of source systems that this data can originate from, beyond core systems like the PSS systems, volumes are very large. The velocity at which this data is generated is also a challenge, as well as data quality issues (missing values, inconsistent data input) and schema variations across systems in the travel ecosystem.</p><p>Of course, this challenge of volume, velocity and variability in the data leads to a very heavy computational load. It scales in a non-linear fashion (a square relationship with data volume), since each record needs to be compared to all other records to check if they could be a match. For instance, if an input dataset has 1000 records, it would take 500,000 comparisons to check for all possible combinations of matches. This doesn’t sound too bad, but if the input data size is 100 million records, this corresponds to 5x10¹⁶ comparisons! To compound this issue, many fields (first name, surname, date of birth, phone, email, etc) for each potential match need to be analysed using fuzzy matching algorithms, which by themselves can be expensive computational operations.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/625/1*r0yjiM7zKPUbCnqfNncXwQ.png" /></figure><p>As an example, the OpenJaw data science team recently implemented our identity resolution algorithm for a large international airline. The ‘cleansed’ input data had approximately 100 million records and we were searching for matches across 7 fields / identifiers. If we were to compare all records for potential matches it would have required 3.5x10¹⁷ comparisons. Obviously, it is not reasonable to compute this, as 10¹⁷ is an unimaginably large number. For comparison, it’s estimated that there are 250 billion (2.5x10¹¹) stars in the Milky Way, this number is 1 million times larger. To help alleviate the data volume issues, we used techniques to reduce the number of potential matches and use distributed computing frameworks.</p><h4><strong>Technology used at OpenJaw for identity resolution</strong></h4><p>Given the Big Data workloads that real-world identity resolution in travel involves — spanning volume, velocity and variety — we built our solution at OpenJaw using Apache Spark.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jvxJNDq39isZ-Kgf9VsSAQ.png" /></figure><p>This platform is particularly well suited to identity resolution because it has capabilities that span ETL (extract, transform, load), machine learning and graph algorithms all in a highly scalable, parallel processing framework. It also allows us to easily deploy on both Cloud and on-premise infrastructure.</p><p>In our Cloud implementation we use Amazon Web Services (AWS). This spins up an Elastic Map Reduce (EMR) cluster that consumes input datasets stored in AWS S3.</p><p>For on-premise solutions, we use a Cloudera cluster running Apache Hadoop YARN as the application master and Hadoop Distributed File System (HDFS) for storage.</p><p><strong>In part 2 of this article…</strong></p><p>In <a href="https://medium.com/the-openjaw-data-science-blog/identity-resolution-for-travel-retailing-part-2-3b5752cd9930">part 2 </a>we will describe in detail the identity resolution algorithm we have developed at OpenJaw that runs at scale on the infrastructure described above. This is a sophisticated algorithm that uses a combination of probabilistic (‘fuzzy’) matching algorithms, graph theory and machine learning to resolve identities across very large (100M+ records) customer data sets.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b16ff6d1edd6" width="1" height="1" alt=""><hr><p><a href="https://medium.com/the-openjaw-data-science-blog/identity-resolution-for-travel-retailing-part-1-b16ff6d1edd6">Identity resolution for travel retailing (part 1)</a> was originally published in <a href="https://medium.com/the-openjaw-data-science-blog">The OpenJaw Data Science Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[In travel retailing, some customers are more equal than others (part 1)]]></title>
            <link>https://medium.com/the-openjaw-data-science-blog/in-travel-retailing-some-customers-are-more-equal-than-others-part-1-202eb78d869f?source=rss-7bd43b8b9097------2</link>
            <guid isPermaLink="false">https://medium.com/p/202eb78d869f</guid>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[segmentation]]></category>
            <category><![CDATA[customer-lifetime-value]]></category>
            <category><![CDATA[clv]]></category>
            <dc:creator><![CDATA[The OpenJaw Data Science Team]]></dc:creator>
            <pubDate>Fri, 01 Nov 2019 09:56:37 GMT</pubDate>
            <atom:updated>2019-11-01T09:56:37.130Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-RC-HmuCN_bYirlxrfzlgA.jpeg" /></figure><p>In this article, we share how we predict Customer Lifetime Value (CLV) at OpenJaw and explain why this metric is so important for travel retailing. This is the first part of a three part article on this topic.</p><p><em>By </em><a href="https://www.linkedin.com/in/john-carney-40328920/"><em>John Carney</em></a><em>, </em><a href="https://www.linkedin.com/in/charles-wang-174970144/"><em>Yuxiao Wang</em></a><em>, </em><a href="https://www.linkedin.com/in/aurenferguson/"><em>Auren Ferguson</em></a><em>, </em><a href="https://www.linkedin.com/in/dr-beibei-flynn-cta-2681129/"><em>Beibei Flynn</em></a><em> and </em><a href="https://www.linkedin.com/in/gavin-kiernan-365a359/"><em>Gavin Kiernan</em></a><em>.</em></p><p><strong>Introduction</strong></p><p>In the travel industry, some travellers are more equal than others. This is evident every time you step on to an aircraft — some passengers sit comfortably sipping champagne in business class awaiting departure, whilst others struggle to cram their carry-on baggage into the overhead bins in economy. But of course, this added luxury comes at a cost; on average, a business class ticket costs four times as much as an economy ticket and sometimes this difference can be much larger, even on the same flight.</p><p>These differences in ticket price translate to large differences in customer value for a travel retailer over time. This is particularly important when profit margin is considered e.g. a very frequent traveller who only purchases the lowest cost economy seats may have negligible value to an airline in profit margin terms, compared to an infrequent traveller who only purchases business or first class tickets.</p><p>In this respect, it is critical that any travel retailer engages with customers in a differentiated, personalised fashion given the dramatic differences that customers have in terms of margin contribution to the business. Despite this reality, travel retailers struggle today to even measure the relative value of their customers.</p><p>There are two primary mechanisms travel retailers use to measure the relative value of their customers in travel today. The first is through points accumulated in a loyalty program and the resultant ‘loyalty tier’ that a customer qualifies for. The second is through the class of flight or accommodation they have booked for a specific trip.</p><p>However, in both cases, these measures are incomplete and in some scenarios completely inaccurate. And it is obvious why — although a loyalty tier does capture a measure of customer value over time, it is a coarse measure that doesn’t differentiate the value of customers inside a tier, or account for the fact that some ‘loyal’ customers don’t participate in the loyalty program at all.</p><p>The fare class for a specific trip is also incomplete, but in a different way — a customer may be flying economy today because she is travelling for leisure, but this customer could in fact be a very high-value customer for the travel retailer overall, as she travels exclusively in business class for work purposes.</p><p>A better way to measure customer value in travel is to quantitatively measure and predict Customer Lifetime Value (CLV). This approach is well established in other areas of consumer retail, especially financial services, telecommunications and online retail. In this article we describe the CLV methodology we have developed at OpenJaw to measure and predict customer value in travel.</p><h4><strong>Background to CLV and its variants</strong></h4><p>Fundamentally, the purpose of CLV is to assess the financial value of each customer. It is different to other customer metrics such as customer profitability, in that it is predictive. A popular, formal definition goes as follows:</p><blockquote><em>CLV is the present value of the future cashflows attributed to the customer during his / her entire relationship with the company. (1)</em></blockquote><p>The predictive model used to estimate CLV can have varying levels of sophistication, ranging from complex machine learning methods to simple heuristics. CLV follows the concept of applying present value to cash flows attributed to the customer relationship. In this respect, it is a measure of the financial value of a customer to the firm, or a theoretical upper bound on the amount a firm should invest to acquire the customer, or to retain the customer.</p><p>CLV methods have traditionally been used in relationship focused businesses such as financial services and telecommunications, where there is also contract with the customer. However, recently, CLV has been applied to other areas of consumer retail that are more transactional in nature e.g. travel and e-commerce. This is mainly because it is very useful in practice. For example, if you want to offer a financial incentive to a customer (e.g. a discount) to increase the probability of conversion, then CLV can be used to determine at what level this discount should be set at, for the discount to make economic sense to the retailer.</p><p>One of the confusing things about CLV is that there are many variants, especially when you read articles that fall outside the rigour of peer-reviewed scholarly publications. This is generally understood amongst experienced data scientists in the field — in fact, there are a few papers that attempt to provide some clarity by identifying all of the variants of CLV and highlighting their differences. A good example of this can be found in (2). In this paper true CLV is compared with:</p><ul><li>Post Acquisition Value (PAV);</li><li>Residual Lifetime Value (RLV);</li><li>Historical Value (HV);</li><li>Sales CLV (S-CLV);</li><li>Finite horizon CLV;</li><li>Un-discounted CLV;</li><li>Cash-flow CLV and</li><li>Accrual-based CLV.</li></ul><p>As you can see, the field of CLV is a minefield, with many variants and a myriad of vague publications online in web forums that confuse one approach over another. In this article we attempt to provide some clarity and describe the methodology we have developed at OpenJaw for estimating CLV in travel, mindful of some of the practical hurdles in this setting.</p><h4><strong>Applying the principles of CLV to travel retailing</strong></h4><p>There are several challenges in estimating CLV in travel retailing. Many of these relate to the fact that there is significant variability in pricing and profitability for the same or similar products across multiple dimensions e.g. there are multiple fare buckets for seats on the same airplane and route, plus these can vary at different times of the day, year and sales channel.</p><p>Similarly, for accommodation, price for the same hotel room can vary by sales channel and time of the year or promotion. And it gets even more complex when products are bundled into packages as they are by tour operators that sell holidays. Dynamic bundles are created with discounts relative to purchasing products individually e.g. flight and hotel, but the bundle is priced opaquely with pricing driven by contracts negotiated with individual suppliers such as hotel chains.</p><blockquote>Therefore, the profit earned on each customer transaction can be highly variable in travel retailing.</blockquote><p>Another important difference in travel retailing, which largely stems from the fact that travel retailing is transactional in nature (versus contractual), is that customer acquisition costs are rarely tracked, especially at the customer level.</p><h4><strong>In part 2 of this article…</strong></h4><p>As you can see, it is not easy to model and predict CLV in the travel domain. In part 2 of this article we will describe how we navigate this minefield of complexity at OpenJaw to generate accurate estimates of CLV.</p><h4><strong>References</strong></h4><p>(1) Farris, Paul W.; Neil T. Bendle; Phillip E. Pfeifer; David J. Reibstein (2010). <em>Marketing Metrics: The Definitive Guide to Measuring Marketing Performance.</em></p><p>(2) McCarthy, Daniel. <em>A general framework for customer lifetime value. </em><a href="https://www.dropbox.com/s/xjak7pezn6i9m06/CLV%20framework.pptx?dl=0">https://www.dropbox.com/s/xjak7pezn6i9m06/CLV%20framework.pptx?dl=0</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=202eb78d869f" width="1" height="1" alt=""><hr><p><a href="https://medium.com/the-openjaw-data-science-blog/in-travel-retailing-some-customers-are-more-equal-than-others-part-1-202eb78d869f">In travel retailing, some customers are more equal than others (part 1)</a> was originally published in <a href="https://medium.com/the-openjaw-data-science-blog">The OpenJaw Data Science Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Welcome to the OpenJaw Data Science Blog!]]></title>
            <link>https://medium.com/@brian_lewis/welcome-to-the-openjaw-data-science-blog-49f353252144?source=rss-7bd43b8b9097------2</link>
            <guid isPermaLink="false">https://medium.com/p/49f353252144</guid>
            <category><![CDATA[segmentation]]></category>
            <category><![CDATA[personalisation]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[customer-data-platform]]></category>
            <category><![CDATA[marketing]]></category>
            <dc:creator><![CDATA[The OpenJaw Data Science Team]]></dc:creator>
            <pubDate>Fri, 01 Nov 2019 09:50:48 GMT</pubDate>
            <atom:updated>2019-11-01T13:27:11.244Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TUFzAUVcq-Y6FGC3q6qC5w.png" /></figure><p>Insights and opinions for data scientists and engineers in the travel industry.</p><p><em>By </em><a href="https://www.linkedin.com/in/john-carney-40328920/"><em>John Carney</em></a><em>, Chief Data Scientist, OpenJaw Technologies</em></p><p><a href="http://www.openjawtech.com/">OpenJaw</a> started its data science journey on 1st May 2017. I remember it well because it was my first day at the company! Since then, we have hired a world class team of data scientists and data engineers, built a petabyte scale Big Data platform for customer analytics and personalisation in travel and deployed this platform into production for some of the largest and most innovative travel brands in the world. We are now, in October 2019, generating insights for personalised travel retailing for over 40 million travellers globally; quite an achievement from a standing start just over 2 years ago.</p><p>The purpose of this blog is to openly share the lessons learned and technical insights we have accumulated over the past 2 years on our journey. We will continue to do this as we develop our platform into the future. We want to give something back to the data science community in travel to accelerate progress in the field and also to engage with the field in an open, pro-active and collaborative way. If you are working as a data scientist in the travel industry, or in the field of personalisation in any industry, it is highly probable that you are currently trying to solve the same or very similar problems as we are here at OpenJaw.</p><p>In our first series of articles, we will share with our readers the methods we use to accurately predict customer lifetime value, perform identity resolution, predict customer purchase propensity and segment customers with machine learning — all ‘pillars’ of personalisation for any online retailer or advertiser today. In the area of data engineering, we will focus on the key topics of data pipeline management, ETL and data warehousing, spanning technologies such as Apache Spark, Hadoop and Airflow. We will also touch on the challenges of data visualisation at scale and how to optimise Big Data deployments in the Cloud.</p><p>More than anything though, this blog is a testament to the amazing technical creativity, determination and hard-work of the data team here at OpenJaw. We have some emerging stars of data science here and I think you will enjoy listening to what they have to say and taking the time to understand how they think and solve the very challenging data science and Big Data engineering problems they face every day.</p><p>Most of our articles are written in consumable parts of about 1,000 words each, but we will endeavour to cross-reference them so that readers can see how everything fits into coherent vision for data science in travel.</p><p>I do hope you enjoy our articles. Feel free to reach out to us with feedback and questions. And don’t forget to clap! Happy reading!</p><p><a href="https://www.linkedin.com/in/john-carney-40328920/"><em>John Carney</em></a><em>, Editor, The OpenJaw Data Science Blog</em></p><p>Data.Science@openjawtech.com</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/457/1*U4Pj7aS5Jx5nMWQauC_MCA.png" /><figcaption><a href="https://www.linkedin.com/in/john-carney-40328920/">Dr. John Carney, Chief Data Scientist, OpenJaw Technologies</a></figcaption></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=49f353252144" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>