<?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 Predictive Analytics Group on Medium]]></title>
        <description><![CDATA[Stories by Predictive Analytics Group on Medium]]></description>
        <link>https://medium.com/@pa_group?source=rss-e016f0cb1b4e------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*o2FUcp1ptZIJCfEM1O9D8A.png</url>
            <title>Stories by Predictive Analytics Group on Medium</title>
            <link>https://medium.com/@pa_group?source=rss-e016f0cb1b4e------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 11:20:29 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@pa_group/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[Ensemble Machine Learning]]></title>
            <link>https://medium.com/autostatplatform/ensemble-machine-learning-bbdc342c936c?source=rss-e016f0cb1b4e------2</link>
            <guid isPermaLink="false">https://medium.com/p/bbdc342c936c</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[analytics]]></category>
            <category><![CDATA[ensemble-learning]]></category>
            <category><![CDATA[gradient-boosting]]></category>
            <dc:creator><![CDATA[Predictive Analytics Group]]></dc:creator>
            <pubDate>Tue, 10 Dec 2019 05:17:00 GMT</pubDate>
            <atom:updated>2019-12-10T06:33:56.296Z</atom:updated>
            <content:encoded><![CDATA[<h4>Using a range of ensemble machine learning frameworks to predict the presence of heart disease.</h4><p>In this practical tutorial, we will undertake exploratory analysis of a dataset that contains records of 303 individuals with varying degrees of symptoms that may be used to indicate the presence of heart disease, as determined by the patient’s angiographic disease status.</p><p>We will then apply ensemble methods including <strong>Adaptive Boosting</strong> and <strong>Gradient Boosting</strong> algorithms to build a model that accurately predicts a patient’s heart disease status based on the attributes contained within the dataset.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*U1n8fZ4vFjDNJIhwy8tUuQ.png" /></figure><h3>Data information</h3><p>Common measurements were taken from each patient including blood pressure, type of chest pain, electrocardiographic abnormalities, resting blood sugar and so on. The original dataset can be found at <a href="https://archive.ics.uci.edu/ml/machine-learning-databases/heart-disease/">https://archive.ics.uci.edu/ml/machine-learning-databases/heart-disease/</a>.</p><p>We use the processed cleveland data which contains 14 attributes of the original dataset. The attributes are:</p><ol><li><strong><em>Age:</em></strong><em> age in years</em></li><li><strong><em>Sex: </em></strong><em>sex (1 = male; 0 = female)</em></li><li><strong><em>Cp: </em></strong><em>Chest pain type (Value 1 = typical angina, Value 2 = atypical angina, Value 3 = non-anginal pain, Value 4 = asymptomatic pain)</em></li><li><strong><em>Trestbps</em></strong><em>: resting blood pressure in mm/Hg on admission to the hospital</em></li><li><strong><em>Chol</em></strong><em>: serum cholestorol in mg/dl</em></li><li><strong><em>Fbs</em></strong><em>: fasting blood sugar &gt; 120 mg/dl (1 = true; 0 = false)</em></li><li><strong><em>Restecg</em></strong><em>: resting electrocardiographic results (Value 0 = normal, <br>Value 1 = having ST-T wave abnormality (T wave inversions and/or ST elevation or depression of &gt; 0.05 mV), Value 2 = showing probable or definite left ventricular hypertrophy by Estes’ criteria)</em></li><li><strong><em>Thalach:</em></strong><em> maximum heart rate achieved</em></li><li><strong><em>Exang:</em></strong><em> exercise induced angina (1 = yes; 0 = no)</em></li><li><strong><em>Oldpeak:</em></strong><em> ST depression induced by exercise relative to rest (refer Restecg)</em></li><li><strong><em>Slope: </em></strong><em>the slope of the peak exercise ST segment (refer Restecg) (Value 1 = upsloping, Value 2 = flat, Value 3 = downsloping)</em></li><li><strong><em>Ca:</em></strong><em> number of major vessels (0–3) colored by flourosopy</em></li><li><strong><em>Thal: </em></strong><em>thalium stress test result (Value 3 = normal, Value 6 = fixed defect, Value 7 = reversable defect)</em></li><li><strong><em>Num (the predictor attribute)</em></strong><em>: diagnosis of heart disease* (angiographic disease status) — (Value equals 0 : &lt; 50% diameter narrowing, Value greater than 0 : &gt; 50% diameter narrowing)<br>*For this tutorial, we will assume that 0 means no heart disease, and 1,2,3,4 means heart disease.</em></li></ol><h3>Data Management, Cleaning and Preprocessing</h3><p>Before proceeding, we need to examine the data to gain more understanding about its underlying properties. This can be completed by heading to the <strong>Data Manager </strong>module in <a href="https://www.autostat.com.au/">AutoStat®</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*b_LSwQGd8HEzINXgtAGLZQ.png" /><figcaption><strong>Fig 1. </strong>AutoStat® Data Upload module.</figcaption></figure><p>Let’s have a look at the data and undertake some basic descriptive statistics.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*srT9SjKq_0f9Q69DmBKOQg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bMFFEGocRnC7MrDM5VRHOQ.png" /><figcaption><strong>Fig 2. </strong>AutoStat® Data Managfer; <strong>Fig 3. </strong><em>Descriptive Statistics</em></figcaption></figure><p><strong>Changing Data Types</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ULk8Lp5Dr9gDQqxiB9FxBg.png" /><figcaption><strong>Fig 4. </strong>Datatype Conversion</figcaption></figure><p>It also appears that some of our numerical variables are actually categorical<em>. </em>In the case of the <em>Thal </em>variable, we have three possible values. Although they present as numbers <em>(3, 6 </em>and <em>7)</em>, we know from the data description that these numbers simply represent three different categories. Let’s change the data type to <em>categorical.</em></p><p><strong>Descriptive Statistics and Missing Values</strong></p><p>After we convert all the variables to the correct data type, we can still see that there are missing values in varibles <em>Ca</em> and <em>Thal</em> from the<strong><em> Descriptive Statistics </em></strong>output.</p><p>As we have explored in our other tutorials, there isn’t a hard-and-fast approach to dealing with missing values. Depending on the situation and underlying data, we may impute a certain value (such as mean of the other values) in place of the missing values; treat them as zero; or remove those observations entirely. In this case, as we only have a few missing values, we’re just going to remove the ones with missing values.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TCRfciMFh_G_kWQz5kjNRA.png" /><figcaption><strong>Fig 5. </strong>Descriptive Statistics &amp; Missing Value imputation</figcaption></figure><p><strong>Bucketing values</strong></p><p>In this scenario, the presence of heart disease is not binary, in that the diagnosis of heart disease is based on a spectrum of outcomes (where 0 = no heart disease, and an increasing severity of heart disease from outcomes 1–4). For this tutorial, we are going to define the values into one of two possible outcomes — they either present with heart disease, or they don’t.</p><p>To do so, we need to create a new variable, that <em>buckets </em>the variables into binary outcomes. The new variable is called <em>Target, </em>and is created by converting 0 values to “<em>No Presence</em>” (No heart disease presentation) and converting <em>1, 2, 3, 4 </em>values to <em>“Presence” </em>(heart disease presentation).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pXdjjsyD3IV5y9CVcU0PoA.png" /><figcaption><strong>Fig 6. </strong>Placing categorical variables into buckets.</figcaption></figure><p><strong>Splitting into train and test</strong></p><p>When you are performing classification using machine learning or statistical models, you may wish to split the data into a <strong>training</strong> and a <strong>test</strong> data set. In this case, we want to have an equal split over the stratum of <em>Target</em>, and we have chosen to use 80% (238) of the data to <em>train</em> the models and the remaining 20% (59) to <em>test</em> the predictive ability of the model.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GcrE-bUx03sBYhhM8pT5rA.png" /><figcaption><strong>Fig 7. </strong>Splitting data into Test and Training sets, stratified over the Target variable.</figcaption></figure><p>Now we have taken care of data housekeeping, let’s undertake some exploratory analysis.</p><h3>Visualising the data</h3><p>Before we undertake any modelling, we should visualise the data and variables so that we can gain an understanding of the relationships between each other, and check for obvious data error issues.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pLDJr4s7uBHaqQBOOLgqHQ.png" /><figcaption><strong>Fig 8.</strong> AutoStat® Visualisation Module</figcaption></figure><p>In this analysis, we are interested in finding out which variables are related to <em>Target</em>. So we’ll project the data using a range of charts that will help us understand the relationships between the variables.</p><p>Let’s explore a few chart types.</p><p><strong>Bar chart</strong> of <em>Target</em> with other categorical variables, breaking the visualisation down into various subgroups:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*n1D-_r5wG-M9cFDJey2NaA.png" /><figcaption><strong>Fig 9. </strong>Bar Chart — Patients with/without heart disease, male vs. female.</figcaption></figure><p><strong>Distribution </strong>of <em>Age</em> with subgroup <em>Target:</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*r5dwiOnOgqfa-KvMR0TL3w.png" /><figcaption><strong>Fig 10. </strong>Distribution plot of patient ages; with and without heart disease.</figcaption></figure><p>In this chart, we can see the distribution of patients with and without heart disease by age. Interestingly, we can see that the age distrubution of people who do not have heart disease follows a normal distribution, whereas the age distrubution of people who <em>do have heart disease</em> is left skewed. This may indicate that elder people are more likely to get heart disease.</p><p><strong>Pairplot </strong>of all numerical variables:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5jjdB-TdZwc4etRdWlCqiw.png" /><figcaption><strong>Fig 11. </strong>Pairplot of all continuous variables.</figcaption></figure><p>The pairplot shows the histogram and density of each numerical variable and also shows the correlation of each pair of variables. From the pairplot we can see that, among others:</p><ul><li><em>Chol</em> and <em>Oldpeak</em> are right skewed, and <em>Thalach</em> is left skewed.</li><li><em>Age</em> has strong positive relationship with <em>Tresbps,</em> and a strong negative relationship with <em>Thalach</em>.</li></ul><p><strong>Scatterplot </strong>of <em>age vs. Thalach (maximum heart rate achieved):</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WqumI50zctMI7mlV42TIEQ.png" /><figcaption><strong>Fig 12. </strong>Scatterplot of <em>Age vs. Thalach, with/without heart disease.</em></figcaption></figure><p>The graph shows that <em>Age</em> and <em>Thalach</em> have strong negative relationship and perhaps unsurprisingly, people are less likely to present with heart disease with low <em>Age</em> and high <em>Thalach</em>.</p><h3><strong>Predictive modelling using ensemble methods</strong></h3><h4>Adapative Boosting (ADA-boost)</h4><p>Adaptive Boosting (AdaBoost) is a supervised learning algorithm that is trained by sequentially forming an additive model of base learners (for example, small Classification and Regression Trees (CARTs)) which may only predict better than a random guess.</p><p>A probability distribution is defined over the training data and updated with the estimation of each new base learner to allow the next learner to focus on training errors made by its predecessor.</p><p>Additionally, coefficients are estimated for each base learner which quantify its overall influence on future predictions following training.</p><p>In other words, AdaBoost combines multiple weak classifiers into a strong classifier. Each weak classifier is a stump which is a simple decision tree of its own, with only one node. The AdaBoost method dictates that the error of the previous classifier is used to train the next classifier. <a href="https://help.autostat.com.au/help/adaboost">For a deeper dive on the mathematics behind Adaptive Boosting, check out our documentation here.</a></p><p>The <strong>Model Builder </strong>module allows you to set up your train and test data, select machine learning model of choice, choose model-specific test options (training sets, cross-validation etc.), drag and drop your target and predictor variables, and of course, choose your model parameters.</p><p>To start with, let’s drag and drop our train/test datasets and pick <strong>AdaBoost.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DS23B7TV3WKCREGpmo-tTw.png" /><figcaption><strong>Fig 13. </strong>AutoStat® Model Builder.</figcaption></figure><p>Now we can see the list of variables in the middle pane. Here, we can pick our target and response variables according to what we want to model. Keep in mind a couple of things:</p><ol><li>The <strong># </strong>symbol indicates that the variable is specified as continuous</li><li>The <strong><em>abc</em> </strong>symbol indicates a variable is categorical</li><li>Variables which are in <strong>black text </strong>indicate there are no missing values. If the variable name is in<strong> red text</strong>, missing values are present.</li></ol><p>Now you need to define your model. To do this, drag your <em>Forecast (outcome) </em>variable into the appropriate space (see below). In our case, this variable is <em>Target</em>. Then drag the <em>explanatory </em>variables into the lower space. In our case, we want to predict <em>Target </em>against all the other variables in the dataset.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ErrZ8POH13xHTaB1p5wJzQ.png" /><figcaption><strong>Fig 14. </strong>Selecting variables and parameters in AutoStat® Model Builder.</figcaption></figure><p>Now we can set the parameters of the AdaBoost algorithm for this model.</p><p><strong>A simple explainer of some Adaptive Boosting parameters:</strong></p><ul><li><em>Number of Estimators</em> means how many stumps (simple decision trees) you want to place in the algorithm.</li><li><em>Algorithm</em> SAMME.R uses the probability estimates to update the additive model, while SAMME (Stagewise Additive Modeling using a Multi-class Exponential loss function) uses the classifications only.</li></ul><p>We have choosen <em>Number of Estimators</em> with 10, <em>Algorithm</em> with <em>SAMME</em>. We can now run the model using the <em>Analyse</em> button at the bottom middle of the screen.</p><p><strong>Results Analysis</strong></p><p>Once you implement the analysis, you are taken to the <strong>Model Output </strong>module. The results page is divided into two tabs:</p><ol><li>Model Evaluation: Model Output, Confusion Matrix, Best Statistical Parameter, Classification report and AUC chart</li><li>SSP: Graphs of variable Importance</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*w2MiSA-pev59H20jK5vaiQ.png" /></figure><p>As we can see with <em>number of estimators </em>set to 10, we get the model accuracy of 88.1356%. The variables with the three highest importance, or <em>contribution to the model performance, </em>are <strong><em>Thal</em></strong><em>, </em><strong><em>Cp </em></strong>and <strong><em>Ca</em></strong><em>. </em>Back to our data dictionary:</p><p><strong><em>Thal: </em></strong><em>thalium stress test result (Value 3 = normal, Value 6 = fixed defect, Value 7 = reversible defect)</em></p><p><strong><em>Cp: </em></strong><em>Chest pain type (Value 1 = typical angina, Value 2 = atypical angina, Value 3 = non-anginal pain, Value 4 = asymptomatic pain)</em></p><p><strong><em>Ca:</em></strong><em> number of major vessels (0–3) colored by flourosopy.</em></p><p>Based on the result of our this example model, we can see that the outcome of the patient’s <em>thalium stress test, type/severity of anginal pain, </em>and the outcome of their <em>blood vessel fluoroscopy </em>have a notable impact on the presence of heart disease.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/941/1*Hgwj7TQKpi8BfGnOahk7kA.png" /><figcaption><strong>Fig 16. </strong>Re-running AdaBoost model with varying estimators.</figcaption></figure><p>Feel free to try a number for <em>Number of Estimators</em> and examine the results:</p><p>It seems like <em>Number of Estimators</em> reach the local minimum around 23. When <em>Number of Estimators</em>&gt;23, the model is overfitted with high variance. When <em>Number of Estimators</em>&lt;23, the model is underfiited woth low variance but high bias.</p><h4>Gradient Boosting (GBM)</h4><p>Gradient boosting is a supervised learning algorithm which sequentially trains a series of base learners (typically weak learners which may have a level of predictive power slightly better than a uniform random guess) to make predictions. Because their functional representation is additive, a large number of learners can be trained to develop a relatively sophisticated model.</p><p>Both Gradient Boosting and AdaBoost create weak learners in sequence. The main difference with the Gradient Boosting algorithm is that the weak classifiers can produce decision trees with more than one node. Additionally, it is a more generic algorithm to find approximate solutions to the additive modeling problem. For more information on Gradient boosting, <a href="https://help.autostat.com.au/help/gbm">check out our documentation here.</a></p><p>Drag and drop the datasets, select <strong>Gradient Boosting </strong>as the model of choice, and drag in the same variables into the appropriate sections as we did with AdaBoost. Now we can set the parameters of the algorithm.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Mu8-gn900-yp3CmEEWOlXw.png" /><figcaption><strong>Fig 17. </strong>Model Builder — Gradient Boosting parameter selection.</figcaption></figure><p><strong>A simple explainer on some Gradient Boosting parameters:</strong></p><ul><li>Gradient boosting algorithms can follow a number of different <em>loss functions. Deviance, </em>for example, is a loss function that &lt;&lt;INSERT HERE &gt;&gt;&gt;</li><li>You can select a number of different <em>learning criteria. Friedman Mean-Squared Error, </em>for example, follows that &lt;&lt; insert here &gt;&gt;</li><li><em>Maximum Tree Depth</em>, <em>Minimum Split</em>, <em>Minimum Leaf Samples</em> and <em>Subsample</em> controls the shape of each tree.</li></ul><p>Varying the number of trees (estimators), tree depth, learning rate, number of splits and leaves will give you varying results. While more advanced parameter tuning practices are beyond the scope of this tutorial, <strong>AutoStat</strong>®<strong> </strong>will automatically select the parameter values which generate you the lowest testing error. This is what we call <em>Bayesian Hyperparameter Optimisation. </em>For more information, <a href="https://towardsdatascience.com/a-conceptual-explanation-of-bayesian-model-based-hyperparameter-optimization-for-machine-learning-b8172278050f">check out this handy Medium article here</a>.</p><p><strong>GBM Results</strong></p><p>After we have ran the GBM model, we can now analyse the results of the model.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*w76DmB2YwMz_a4LvzQ6EgA.png" /></figure><p>As you can see from the <em>Best Statistical Paramenter</em>, the best maximum tree depth is 5, the best minimum samples split is 3, the best minimum samples leaf is 3 and the best learning rate is 0.0743. The accuracy we get is 88.1356% which is slightly lower than AdaBoost.</p><h3>Building a data science pipeline</h3><p>A critical element of data science is the ability to turn your project into an end-to-end automated workflow that re-runs either on-demand, or as new data becomes available. Pipelines are automated workflows that allow the organisation, decision-maker or researcher to productionise their prototypes and integrate them with existing workflows.</p><p>While we might not need to see the <em>Cleveland Heart Disease</em> project updated on-demand or as new data becomes available<em> (alas, the dataset was published 30 years ago!)</em>, this is how our project might look as a production-ready pipeline.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mQCmeTyPOcs3Ga-bbWjGww.png" /><figcaption><strong>Fig 19. </strong>AutoStat® Pipeline Module.</figcaption></figure><p>In the next tutorial, we will explore how to build and deploy data science pipelines so that any data project can be productionised from end-to-end, from accessing the data itself, through to cleaning and preprocessing; to visualisations, model specification and finally, dissemination of results via APIs (to be consumed elsewhere), user-friendly dashboards, and publishable reports, all without writing a line of code.</p><h3>See for yourself</h3><p>Want to start running your own machine learning projects?</p><p>Head to our website for a free trial here:</p><p><a href="https://www.autostat.com.au/">https://autostat.com.au</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bbdc342c936c" width="1" height="1" alt=""><hr><p><a href="https://medium.com/autostatplatform/ensemble-machine-learning-bbdc342c936c">Ensemble Machine Learning</a> was originally published in <a href="https://medium.com/autostatplatform">autostatplatform</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Mixed Effects Models]]></title>
            <link>https://medium.com/autostatplatform/mixed-effects-models-faf1ef306633?source=rss-e016f0cb1b4e------2</link>
            <guid isPermaLink="false">https://medium.com/p/faf1ef306633</guid>
            <category><![CDATA[bayesian-statistics]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[statistics]]></category>
            <dc:creator><![CDATA[Predictive Analytics Group]]></dc:creator>
            <pubDate>Tue, 19 Nov 2019 07:28:27 GMT</pubDate>
            <atom:updated>2019-11-19T07:28:27.077Z</atom:updated>
            <content:encoded><![CDATA[<h4>Analysing chick weights over time under four different diet categories using AutoStat®</h4><p>In this tutorial we will use a linear mixed effects model and compare it to a standard linear regression.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*n6vCG2Cufitk8DKwR8FXTA.jpeg" /><figcaption>Modelling chick weights over time. Credit: Newfabrika — Freepik.com</figcaption></figure><p>The usual representation of the linear regression is given as:</p><p><strong><em>Y=Xβ+ϵ</em></strong></p><p>where <strong><em>Y </em></strong>is the outcome measure, <strong><em>X </em></strong>is a design matrix of explanatory variables, <strong><em>β </em></strong>are the estimated coefficients of the model and the residual (<strong><em>error</em></strong>) term is represented as <strong><em>ϵ.</em></strong></p><p>The assumptions of a linear regression include:</p><ul><li>The data are independent samples</li><li>The errors are distributed as N(0,σ²)</li></ul><p>We now consider a situation where data are taken longitudinally as a growth study. For such data, there are repeated measures of individuals, and often the variance around the growth profiles differs over time (<em>not heterogeneous</em>).</p><p>Conceptually, it is easy to understand why heterogeneity of errors occurs in growth profiles. Imagine a situation of human children being monitored, from birth, for weight over a 5 year period. When we are born, there is only a small variance in birth weights, as by necessity, we are small when we are born. From birth, each individual will follow their own individual “growth curve”, due to factors such as genetics, environment and diet. Our growth is not exactly on-target with such a profile, and we may vary either side of our individual “mean growth” at any time. Additionally, while we may start life with a variation of say 10% between weights, as we grow, a 10% difference in weights will also increase in range, therefore, it would not be possible to maintain the same range of values as we grow.</p><blockquote><strong>In this study, we will look at the growth of poultry chicks over time. These chicks are raised in a similar environment, but are subject to 4 different diets, with diet allocation being assigned randomly at birth.</strong></blockquote><h3>Data Visualisation</h3><p>Here is the mean profile of Weight by diet over time. This profile looks like a potential candidate for a spline on the time variable, with growth rates increasing at around the 12 weeks of age, and possibly also at 6 weeks.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*R0l7QgoFQbJapaTkrQIP1A.png" /><figcaption><strong>Fig 1. </strong>AutoStat® Visualisation module. Line chart of chick weights over time by diet type.</figcaption></figure><p>If you do a scatterplot, instead of the mean profile, you can immediately see that the variance around a fitted profile will not be constant, with variance increasing as the chicks grow.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IgIZFWrIVTNS5ABiFs5r0A.png" /><figcaption><strong>Fig 2. </strong>Scatterplot of chick weights over time. Note the variance in weights across diet groups increasing over time.</figcaption></figure><p>From the plot we note that <strong>Diet 1</strong> seems to have the lowest weights for the chicks, but the other diets are not clear as to whether they are diffferent. We also notice a particular chick in <strong>Diet 2</strong> (purple dots) that has very low body weight compared to the other chicks in its treatment group.</p><p>In this particular set of measurements, not all chicks are measured at each time. A quick way to check this is using <strong>sub-groups</strong> in the <strong>Visualisation</strong> module. Here is an example plot, and we note feature like chick 18 dropping out of the study early, chick 15 not thriving and dropping out halfway through the study, and chick 21 growing quite quickly.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vvtVVuoBsirseos9_nK6BQ.png" /><figcaption><strong>Fig 3. </strong>Line plot of individual chick weights over time. Note the thriving growth from Chick #1, #2, #14 and #21 versus Chick #15, #16, #18.</figcaption></figure><p>Here is the same plot, but placed on the 4 separate charts for diet responses. We do this using by plotting each <em>Chick ID</em> as a <strong>sub-group, </strong>and each <em>Diet</em> category in the <strong>sub-plot:</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1dGh9AY3SzbUuamSclng_w.png" /><figcaption><strong>Fig 4. </strong>Individual chick weights over time, by diet category.</figcaption></figure><h3>Linear Regression</h3><p>We firstly conducted a linear regression using a Bayesian approach with variable selection using a spike-slab G-prior.</p><p><a href="https://medium.com/autostatplatform/an-intuitive-guide-to-linear-regression-a90331ba8b8c"><em>Need to brush up on </em><strong><em>linear regression with Bayesian variable selection</em></strong><em>? In our last tutorial, we explored three different Linear Regression approaches: Frequentist, Lasso and Bayesian Variable Selection. You can access this here.</em></a></p><p>The variable selection aspect of this model reflects what we would intuitively think should occur. None of the main effects for <em>diet</em> are shown to have an inclusion probability above 0.15. This makes sense, as the chicks at birth have not been subjected to the diet treatments, and are yet to be randomly assigned to the treatment. If the random assignment is sucessful, these “intercept” terms should not contribute to the model in any significant way.</p><p>Therefore we have the constant term, which in this case is essentially the mean birth weight, an effect of time, as the chicks grow, and an effect of each diet over the standard diet through time. This model only compares each diet to the base diet (Diet 1), not to each other.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tRl0OaUYX7XBus3wo2YeAw.png" /><figcaption><strong>Fig 5. </strong>AutoStat®’s Model Builder- Select Dataset(s), Model, Variables and Parameters before analysing.</figcaption></figure><p>The posterior distributions of the coefficients (below) have been produced using a model average approach. In summary, each coefficient is a product of the coefficient values it takes over each of the models, and weighted on how likely the model is. Therefore, terms that are in the most probable models are given more weight than those which occur in few or less probable models. Terms which appear in only improbable models tend to be shrunk towards zero (0).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/747/1*5ujlio4GtFpLbQM4GX5m6g.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/393/1*Y9aIbt6ni7cHiG3VNdtDNQ.png" /><figcaption><strong>Fig 6. </strong>Linear regression model output.</figcaption></figure><p>We see that the main effects of diet are generally small in magnitude, with the exception of<em> Diet 3</em>, however, this term has large standard deviation, so it is not significant.</p><p>The coefficients for <em>Time </em>(and interactions with <em>Diet</em>) are more substantive, with low standard deviation. We see that the coefficient for time indicates that for each week under <em>Diet 1</em>, chicks would gain, on average, 7 grams in weight, and under the other diet regimes, this gain would be greater.</p><p>We also note that the coefficients are not sensible in several aspects. For example,</p><ul><li>The constant does not reflect the weight of chicks at birth, with it underestimating the birth weight by around 10 grams.</li><li>The estimated sigma is very large, and not reflective of chicks when they are young.</li></ul><p>The first issue would be improved by including a spline term for <em>Time</em>, but the issue with the mean-variance relationship will not be addressed by using the linear regression framework.</p><p>The <em>Actual vs. Prediction </em>plot shows the under-estimation of the weight at birth. You can find this in the <strong>diagnostics</strong> tab in the <strong>Results</strong> module.</p><p>We can also see the large range of the residuals in the model via the histogram in the diagnostics panel, with some residuals in the order of ±100 grams.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/786/1*2e745_9kAmW8VFxROMtX6A.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/787/1*4MkxEZc63O9tjnYYHnsSiQ.png" /><figcaption><strong>Fig 7. </strong>Regression Model<strong> </strong>Predictions vs. Actual chick weights. <strong>Fig 8. </strong>Distribution of model residuals</figcaption></figure><h3>Mixed Models</h3><p>The linear mixed effects model is represented as:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/292/1*wU2CtHwz_OgKmZZiTQenFg.png" /></figure><p>Here, <em>i</em> is an individual chick, <em>Xi</em> is the design matrix of the fixed effects for chick <em>i (</em>as we are familiar with from the standard linear regression), <em>β</em> are the coefficients for the fixed effects and the error term <em>ϵ_i</em> is distributed as N(0,σ²).</p><p>The big change is the addition of a second design matrix <em>Zi</em> and the subject specific random effects <em>b_i</em>. These <em>b_i</em> allow us to model the subject specific means and enable us to capture the marginal dependence among the observations on the <em>i-</em>th unit.</p><p>The distribution of the random effects is <em>b ∼ N(0,D)</em>. The extra variance component <em>D</em> is considered independent of the residual variance (<em>σ²</em>). The analyst has some options for modelling the elements of D. They can be independent between <em>b</em>’s, or there can be a variance-covariance structure between some or all of the <em>b</em>’s.</p><p>Lets start by doing a linear mixed model with a<strong> spline regression</strong> for Time. The model will be:</p><blockquote>Weight =</blockquote><blockquote>Constant <em>or </em>Intercept + Diet + Time + Diet:Time + ChickID + ChickID:Time</blockquote><p>where terms involving Chick are considered to be random.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NUBICWORqnvJOux-29QFNg.png" /><figcaption><strong>Fig 9. </strong>Model Builder — Selecting Linear Mixed Model, outlining fixed and random effects.</figcaption></figure><p>We can see from the output, that many improvements in fit have been made with this model. For example, the Constant is now at a much more reasonable estimate, around 10 grams heigher than the linear regression. This is due to using a spline term in time, not necessarily due to mixed effects modelling. We can see the effect of the spline term by viewing the <strong>variable impact charts</strong> tab:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/790/1*WWNeXRkXEYDHI3OpLrgfCQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/784/1*IHMamjjhiMIu9wwtuWEzEQ.png" /><figcaption><strong>Fig 9. </strong>Variable impact of Time and Time:Diet on model output.</figcaption></figure><p>We see that growth is linear until around 6–8 weeks, and then the slope of the growth changes (but remains fairly linear). The variable impact charts also shed light on the effect of diet over time, with each diet type indicated as having an additional effect on growth (chickens weighing more over time), with <em>Diet 3</em> having a very strong effect on final growth weight.</p><p>Looking at the <strong>standard output</strong> tab we notice that the model is probably subject to overfitting (see below). Biologically, we know that chicks are allocated randomly to the Diet groups, so the terms for Diet are probably not contributing to the model in a meaningful way.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gPGObN0DKSj2-6TY4BZLNg.png" /><figcaption><strong>Fig 10. </strong>Mixed Model standard output</figcaption></figure><p>Looking at the <strong>Diagnostics Tab</strong> we now see residuals that are much improved, with a range now reduced to ±20 grams, and the <em>actual vs predicted</em> graph no longer shows the increasing mean-variance relationship of the linear model.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/780/1*lOe0HO5MKOlXMlZ09arPRg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/780/1*U32Pp__9xEYpzojVeC_zEg.png" /><figcaption><strong>Fig 11. </strong>Regression Model<strong> </strong>Predictions vs. Actual chick weights. <strong>Fig 12. </strong>Distribution of model residuals</figcaption></figure><h4>Model averaging and overfitting</h4><p>To try and reduce the possibility of overfitting, we will now use a <em>Normal spike slab prior </em>on the fixed effects. AutoStat® currently provides both a Normal spike slab and a G-prior spike slab for variable selection. Generally, these models will give similar inferences, particularly when using value priors for the “slab” component of the model.</p><p>The results given in the Stochastic search panel indicate that the main effect of Diet (birth weight for each diet) is very much reduced, with inclusion probabilities of 5% or less.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DuYdqH66wue4YKWEpoGVmg.png" /><figcaption><strong>Fig 13. Results Interpretation: </strong>Stochastic search panel.</figcaption></figure><p>The first model in the Table has a model probability of 0.83, and includes all of the spline terms, and none of the main effect Diet coefficients. The high model probability indicates that the coefficients returned for main effects diet will have shrinkage towards zero (0).</p><p>Here is the table of regression coefficients determined using model averaging, where we can see that these terms for <em>Diet </em>have been subjected to shrinkage, and now are of much lower magnitude than in the previous model.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EnBRGs1r-8WIR1Ik13onnw.png" /><figcaption><strong>Fig 14. </strong>Mixed Model (Normal Spike Slab) — Note shrinkage of Diet variables combatting overfitting.</figcaption></figure><h3>Summary</h3><p>In this tutorial we have explored using splines for non-linear effects in linear regression, and the use of mixed models for repeated measures growth profiles. We then used variable selection to shrink non-required terms in the model.</p><p>We found that:</p><ul><li>Spline terms on non-linear response resulted in better fits, particularly at the intercept;</li><li>Mixed effects models resulted in improved residual variances, as they accounted for variance attributable;</li><li>Variable selction using model averaging resulted in unimportant terms having shrinkage applied to their estimators.</li></ul><h3>See for yourself</h3><p>Want to start running your own statistics and data science projects?</p><p>Head to our website for a free trial here:</p><p><a href="https://autostat.com.au/">Https://autostat.com.au</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=faf1ef306633" width="1" height="1" alt=""><hr><p><a href="https://medium.com/autostatplatform/mixed-effects-models-faf1ef306633">Mixed Effects Models</a> was originally published in <a href="https://medium.com/autostatplatform">autostatplatform</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Using Machine Learning to Predict Citywide Site Developments]]></title>
            <link>https://medium.com/@pa_group/using-machine-learning-to-predict-citywide-site-developments-3916894810ac?source=rss-e016f0cb1b4e------2</link>
            <guid isPermaLink="false">https://medium.com/p/3916894810ac</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[gradient-boosting]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[predictive-analytics]]></category>
            <category><![CDATA[strategy]]></category>
            <dc:creator><![CDATA[Predictive Analytics Group]]></dc:creator>
            <pubDate>Mon, 18 Nov 2019 01:01:21 GMT</pubDate>
            <atom:updated>2019-11-18T01:01:21.191Z</atom:updated>
            <content:encoded><![CDATA[<p><em>A Predictive Analytics Group case study on applying a combination of ML and Statistical frameworks to determine the probability of site development and redevelopment to support long-term urban planning initiatives.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ndYCiJiIlnfi1toTNYo2Zw.png" /></figure><h3><strong>Challenge — assisting with strategic planning</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cxTnxoxqjiarTWA_bU3BUw.png" /></figure><p>City of Melbourne (CoM) is the responsible local government for several inner Melbourne suburbs including Melbourne, Docklands and Southbank. To assist with strategic planning across the precinct, CoM engaged Predictive Analytics Group (PAG) to develop a predictive model of property site development likelihoods.</p><p>Development may be classiﬁed as Residential or Non-Residential. As the two are not mutually exclusive a third target variable labelled, ‘Developed’ was also considered.</p><h3><strong>Solution — developing a machine learning framework</strong></h3><p>PAG developed the model using the data provided by CoM, which is comprised a total of approx. 13,500 observations. Explanatory variables used in the model related to:</p><p>• site value;</p><p>• space (including total site space and unused space);</p><p>• site spatial proﬁle (obtained through shapeﬁle polygons);</p><p>• heritage designation;</p><p>• socio-economic information;</p><p>• the number of dwellings on a site; and</p><p>• other variables capturing diﬀerent aspects of the site and its surrounding area.</p><p>Three distinct binary response variables were modeled as part of the project, namely:</p><p>• residential development;</p><p>• non-residential development; and</p><p>• development.</p><p>A value of 1 for each of these variables indicated the site has been developed in the past while a value of 0 indicated the site had not been developed.</p><p>PAG deployed the following method to estimate the likelihoods:</p><p>1. Feature Engineering.</p><p>2. Train-Validation-Test Split.</p><p>3. Model Development, Evaluation and Iteration.</p><p>4. API Embedding and Development.</p><h3><strong>Outcome</strong></h3><p>The following conclusions (with respect to modelling site development in the CoM precinct) were drawn:</p><p>• It is possible develop a strong predictive model of the likelihood of site development within the CoM precinct;</p><p>• Strong predictive performance can be achieved with models that possess mechanisms for interpretation such as, tree based frameworks including Light Gradient Boosting and Random Forest;</p><p>• Of the three target variables, residential development could be considered the most diﬃcult to predict; and</p><p>• The organisation, variation in variables and overall quality of the MCC’s data was critical to the success of the project.</p><p>The success of the project will assist City of Melbourne to model the likelihood of site development across the city with greater accuracy and enhance the organisations strategic urban planning initiatives.</p><h3>The Platform</h3><p>The PAG data science team used it’s very own proprietaty software, <strong>AutoStat</strong>® to undertake advanced feature engineering and rapid prototyping of machine learning models and parameter tuning without writing any code.</p><p>Once exploratory analysis and model prototyping/tuning had been undertaken within AutoStat®, we developed bespoke APIs in Python for City of Melbourne to integrate the solution into its existing analytics suite, allowing them to re-run the model as new data becomes available.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3rXkitN1mqFpwhidhVjFGA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lMGDefkg3A3f8u3YBq7sOg.png" /></figure><h4>Do it yourself!</h4><p>Thanks to AutoStat®, you no longer need to be a PhD-qualified data scientist or Python/R guru to solve prediction, forecasting and optimisation problems that arise in your organisation. Just plug in your data and get started!</p><p>Head to <a href="https://autostat.com.au/">https://autostat.com.au</a> for more information and start your own data science journey today.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3916894810ac" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[An intuitive guide to Linear Regression]]></title>
            <link>https://medium.com/autostatplatform/an-intuitive-guide-to-linear-regression-a90331ba8b8c?source=rss-e016f0cb1b4e------2</link>
            <guid isPermaLink="false">https://medium.com/p/a90331ba8b8c</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[bayesian-statistics]]></category>
            <category><![CDATA[statistics]]></category>
            <category><![CDATA[linear-regression]]></category>
            <dc:creator><![CDATA[Predictive Analytics Group]]></dc:creator>
            <pubDate>Fri, 15 Nov 2019 04:24:05 GMT</pubDate>
            <atom:updated>2019-11-15T04:34:00.458Z</atom:updated>
            <content:encoded><![CDATA[<h4>Frequentist, Lasso and Bayesian Variable Selection using AutoStat®</h4><p>Our objective to use linear regression to examine the alcohol content (%) of wine based on the cultivar of grape and additional chemical properties.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/948/1*rUWSgvnkOqXOrEFUjWJPbw.png" /></figure><p>The data in this case study are measurements of wine characteristics over <strong>3 different cultivars of grape. The wines are produced in </strong>the <strong>same region of Italy</strong>. The data is available at <a href="https://archive.ics.uci.edu/ml/datasets/wine">https://archive.ics.uci.edu/ml/datasets/wine</a> and contains measurements on 13 different constituents of grapes, namely:</p><ol><li>Alcohol</li><li>Malic acid</li><li>Ash</li><li>Alcalinity of ash</li><li>Magnesium</li><li>Total phenols</li><li>Flavanoids</li><li>Nonflavanoid phenols</li><li>Proanthocyanins</li><li>Color intensity</li><li>Hue</li><li>OD280/OD315 of diluted wines</li><li>Proline</li><li>Cultivar (replacing numeric representation in original data source)</li></ol><h3><strong>Data Management and Data Splitting</strong></h3><p>The dataset is uploaded to AutoStat® via the <strong>Data Manager</strong> module. We proceed by reviewing the data and data types for any anomalies. We engage with the options in the AutoStat® data manager module and examine the data types and review the descriptive statistics.</p><p>Typically, when performing classification using Machine Learning; or predicting using statistical models, the data is split into a <strong>training</strong> and a <strong>test</strong> set. To facilitate our analysis, we split the data equally over the stratum of <em>Cultivar</em>, and select 80% (142) of the data to <em>train</em> the models with the remaining 20% (36) left to <em>test</em> the predictive ability of the model.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AvZTtMCPDnvf-q977KLwFw.png" /><figcaption><strong>Fig 1. </strong>Splitting dataset into test and train, stratified by the Cultivar variable.</figcaption></figure><p>Although we are using a train and test data set in this example, we note that the data set is both small (178 records) and noisy (see visualisations below). Given this, from a data analytic point of view, we do not expect to get good predictions from any model we use, and are conducting the analysis to gain an understanding of which chemical characteristics and cultivars influence alcohol content in wine.</p><h3><strong>Visualising the data</strong></h3><p>Our objective is to examine which chemical measures are related to alcohol content in wine. We begin with a boxplot and add <strong>subgroups</strong> to compare the cultivars.</p><p>A comparison of the subgroups shows that Grignolini cultivar produces a wine that is lower in alcohol content than Barbera or Barolo. The <strong>boxplots </strong>below indicate that the data is not symmetrical about the median value. <strong>Density plots</strong> provide clarity regarding the spread of the population.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0BuaIlWTQWOvJJmNqcrh9w.png" /><figcaption><strong>Fig 2</strong>. Boxplot of alcohol content across the three wine cultivars. <strong><em>Fig 3</em></strong>. Distribution of alcohol content across the three cultivars.</figcaption></figure><p>The other explanatory variables are all <em>continuous</em>. As such, we construct a <strong>scatterplot </strong>to examine any potential relationships.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zn2Hwc51i34Yz5fDsDqYHw.png" /><figcaption><strong>Fig 4. </strong>Scatterplot of alcohol content vs. proline measures.</figcaption></figure><p>It appears there is some linear relationship (albeit noisy), with alcohol content rising as Proline measures increase. The level of alcohol tends to plateau around 15%. This makes biological sense, as most wines we purchase are below 15%. <em>In a follow up tutorial, we will examine the use of splines and transformations in linear regression to better model the biological realities of data.</em></p><p>The pair plot below compares three exploratory variables. This plot is useful in understanding the relationship between the outcome (alcohol content) and the measures we hypothesise may be related to the outcome.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/938/1*-BZiDP-lNLKcjE8z6CiNGw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/910/1*7KLkXs8RswYtVrJeacfRRw.png" /><figcaption><strong>Fig 5. </strong>Pairplot of alcohol content vs. proline vs. color intensity. <strong>Fig 6. </strong>Scatterplot of color intensity vs. proline, higlighting clusters by each cultivar type.</figcaption></figure><p>From the graph, we notice that the relationship between Proline and Colour Intensity looks like there may be clusters, rather than a straight linear response.</p><p>From the scatter plot above, we see that Barolo has generally much higher proline levels then the other 2 wines, whereas Barbera and Grignolino have similar proline levels, but vary in colour intensity.</p><h3><strong>Regression Models</strong></h3><p>We conduct a series of linear regressions, including:</p><ul><li>Standard frequentist.</li><li>Variable selection with Frequentist (Lasso) approach</li><li>Bayesian (G-prior spike-slab) approaches.</li></ul><p>We include all the chemical properties and cultivar as a <em>main effects</em> model only. We do not include any interactions between the variables. The model can be formally expressed follows:</p><p><em>Y=Xβ+ϵY=Xβ+ϵ,</em></p><p><em>where ϵ∼N(0,σ2)ϵ∼N(0,σ2).</em></p><p>The process of specifying the model in AutoStat® is simple and intuitive. Having selected our model framework and dataset we assign our <em>Forecast (outcome) variable</em>, i.e. alcohol, and <em>explanatory variables</em> , i.e. all other variables in the dataset.</p><p>AutoStat® allows for the selection of either the Frequentist or Bayesian modelling paradigms. We select <strong>Frequentist</strong> with <em>No Regularisation</em>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4siVzt3mOrtGlXT8BbdhSg.png" /><figcaption><strong>Fig 7. </strong>AutoStat’s Model Builder Module. Select Dataset(s), Model, Variables and Parameters before analysing.</figcaption></figure><p>The summary table of coefficients indicates that there is a relationship between some of the variables. People often make this assessment based on the <em>p-value</em> of the coefficient, which is an indicator of whether the value is significantly different to zero. The p-value is at the centre of the current debate on reproducibility in science.</p><p>In general, it is reasonable to look at the p-value, but not use it as an absolute bright-line decision making value. Below we consider more robust methods for variable assessment. Based on p-values, we may consider Cultivar and colour intensity are the variables most worth considering in terms of changes in alcohol content.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5uIE5oBjO0mvOTTybuxLjQ.png" /><figcaption><strong>Fig 8. </strong>Frequentist regression model output. Variables with lowest p-values include Color_intensity and Cultivar_Barolo</figcaption></figure><p>The coefficient for Barolo is 0.7246, indicating that the average alcohol content of this wine is higher than the <strong>base level</strong>, i.e. Barbera. The coefficient for Colour intensity is small (positive) in magnitude. How do we interpret this coefficient? An Increasing of 1 proline unit generates an increase in the alcohol content by 0.1243%. The <em>variable impact charts</em> provide a quick view of the change in alcohol content attributed to colour intensity. We see that the change in alcohol (%) being attributed to a change from 2 units of colour intensity to 13 units is approximately 1.5%.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/705/1*YVsUcogNr656BtUP1VNMNA.png" /><figcaption><strong>Fig 9. </strong>Variable impact chart, highlighting relative contribution of color intensity in determining alcohol content.</figcaption></figure><h3><strong>Variable selection</strong></h3><p>It is recognised that including too many variables in the model can reduce the effectiveness of the procedure.</p><p>In the past, techniques such as <em>stepwise</em> where terms were either added or subtracted one at a time from the model, were very common. However, these techniques are now widely discouraged due to their frequent misuse and have been replaced by more efficient algorithms.</p><h4><strong>Lasso regression</strong></h4><p>One option for reducing the number of variables in the model is known as <em>shrinkage</em>. This method places a penalty in the minimization process. This model fits all the required explanatory variables and <em>shrinks</em> them towards zero.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/358/1*31tGSGAzOQpZgeKep98OTQ.png" /><figcaption><strong>Fig 10. </strong>Lasso model output. Note the impact of shrinkage, reducing included variables from 15 down to 9 in the final model.</figcaption></figure><p>Using the default parameters in AutoStat®, the Lasso model reduces the original 15 variables to 9. The non-zero coefficients are displayed in the table below. This is quite different from the p-value interpretation of 2 variables being “significant”. We can see that Colour intensity and the cultivar Barolo are still included in the model, and that the value of these coefficients has been slightly reduced due to the shrinkage the algorithm is designed to achieve. Examining the Table, we see all remaining terms in the model have been reduced in magnitude (except the constant).</p><h4><strong>Bayesian G-prior spike slab models</strong></h4><p>For explanatory power, we estimate the <strong>most probable models</strong> using a <strong>G-prior spike slab algorithm</strong>. The unknown parameters (<em>β</em> , <em>σ</em>) require a prior distribution to be specified to estimate their respective posterior distributions.</p><p>The G-prior provides an estimate of the parameter’s likely location, without requiring the specification of a correlation structure between regressors. This makes it a useful prior for model comparison. The default in AutoStat® is to set <em>g</em> equal to our sample size. Hence, the posterior distribution becomes data driven. The hyper-prior ~β~ is set at zero.</p><p>It may be the case that variables in the full model exhibit collinearity <em>(see pairplot for example)</em>, making assessment of each variable’s contribution to the outcome problematic. Variable selection for dimensionality reduction, and then averaging over probable models, is a common technique for approaching this issue. The need to assess 2^k competing models, requires a strategy to traverse the space. AutoStat® uses stochastic search variable selection, where each of the 2^k models (Mγ) is associated with a binary vector <em>γ</em> where <em>γj</em> = 1 when βj is in the model and 0 otherwise.</p><p>As the model probability, (Mγ), is now another parameter to be estimated, we assign the prior as a uniform distribution for all models. The stochastic search algorithm of Marin and Robert (2007) is then used to determine the posterior probability of each model.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Xjx3bLnZcYWNVuYNM-HNsQ.png" /><figcaption><strong>Fig 11. </strong>Stochastic Search Panel, outlining the posterior probabilities of each variable being included in the model.</figcaption></figure><blockquote><strong>We can see from the Model Probabilities in the stochastic search panel that, in this case, we do not have a clear “winner” for the most probable model.</strong></blockquote><p>Although intuitively, practioners may expect this, and novice users may be disappointed, this is often perfectly acceptable. It is just saying that in this case study, there are several models that fit the data equally well.</p><p>However, when we delve into the coefficients, we see that the parameter estimates are generally smaller in magnitude than the standard linear regression estimates, as was the case in the Lasso approach.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/836/1*p64PzIB3I6BhtZpcH115zg.png" /><figcaption><strong>Fig 12. </strong>Bayesian Variable Selection — Standard output.</figcaption></figure><h3><strong>Results Analysis &amp; Comparison</strong></h3><p>So after running the three models (Standard Frequentist, Frequentist with Lasso variable selection, and Bayesian variable selection with a G-prior spike slab), which model has the greatest predictive power, and why?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Lyv_Nrb1agH8V797jJTs8g.png" /><figcaption><strong>Fig 13. </strong>Results Comparison</figcaption></figure><p>Looking at the prediction metrics of all three models, it is clear to see that the Standard Frequentist regression model performed with the most accuracy, with <em>Mean Absolute; Mean Squared;</em> and<em> Median Absolute Error </em>values notably lower than the Lasso and Bayesian counterparts.</p><h4>Comparison — Lasso vs. Bayesian vs. Frequentist</h4><p>At this point, a practitioner could fairly ask <strong><em>“Why is this so? Shouldn’t the more sophisticated techniques produce better predictions?</em></strong><em>”</em></p><p>Looking at the pairplot of actual alcohol content vs the predictions from the 3 models, we notice something very interesting with the Lasso and the G-prior. There is a <strong>gap</strong> in the predictions that is <strong>filled</strong> in the standard regression prediction.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7617GrN-a0r8L3pRGKQrKQ.png" /><figcaption><strong>Fig 14. </strong>Pairplot of actual alcohol content vs. predicted alcohol content across the three models.</figcaption></figure><p>This is telling the analyst that although we have determined the variables which have the most influence on alcohol content (namely cultivar and colour intensity) either a) we haven’t accounted for an important variable in producing predictions within a certain range, or b) the small size of our dataset and the noise within has caused our test / train data to be unsatisfactory in it’s representation of the underlying population.</p><blockquote><em>In terms of gaining insight, using the combination of modelling approaches enhances our understanding of the different chemical properties, and the affect they have on alcohol content, and it aids in making informed decisions in regards to the direction of future research by embracing uncertainty rather than a bright line rule.</em></blockquote><h3><strong>See for yourself</strong></h3><p>Want to start running your own statistics and data science projects?</p><p>Head to our website for a free trial here:</p><p><a href="https://autostat.com.au/">Https://autostat.com.au</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a90331ba8b8c" width="1" height="1" alt=""><hr><p><a href="https://medium.com/autostatplatform/an-intuitive-guide-to-linear-regression-a90331ba8b8c">An intuitive guide to Linear Regression</a> was originally published in <a href="https://medium.com/autostatplatform">autostatplatform</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to extract insights from big data]]></title>
            <link>https://medium.com/predictive-analytics-group/how-to-extract-insights-from-big-data-69e546d0af15?source=rss-e016f0cb1b4e------2</link>
            <guid isPermaLink="false">https://medium.com/p/69e546d0af15</guid>
            <category><![CDATA[predictive-analytics]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[big-data]]></category>
            <dc:creator><![CDATA[Predictive Analytics Group]]></dc:creator>
            <pubDate>Thu, 31 Jan 2019 03:40:05 GMT</pubDate>
            <atom:updated>2019-01-31T03:40:05.123Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qatotphqCyOKDuGbDl4Qwg.jpeg" /></figure><p>Dr Theo Gazos, Managing Director of Predictive Analytics Group, will be presenting at the 2016 Hadoop Summit in Melbourne on 1 September 2016. His presentation will discuss how to extract insights from big data using predictive analytics with machine learning and artificial intelligence. It will consider issues including:</p><ol><li>How sophisticated analytics can combine structured, unstructured and semi-structured data to develop insights</li><li>Technological developments such as Hadoop, cloud storage and NoSQL, and their implications for sophisticated algorithm development</li><li>Predictive Tools: Conventional Econometrics vs Machine Learning</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=69e546d0af15" width="1" height="1" alt=""><hr><p><a href="https://medium.com/predictive-analytics-group/how-to-extract-insights-from-big-data-69e546d0af15">How to extract insights from big data</a> was originally published in <a href="https://medium.com/predictive-analytics-group">Predictive Analytics Group</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Big data — common pitfalls]]></title>
            <link>https://medium.com/predictive-analytics-group/big-data-common-pitfalls-1416d8fa8668?source=rss-e016f0cb1b4e------2</link>
            <guid isPermaLink="false">https://medium.com/p/1416d8fa8668</guid>
            <category><![CDATA[analysis]]></category>
            <category><![CDATA[predictive-analytics]]></category>
            <category><![CDATA[big-data]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Predictive Analytics Group]]></dc:creator>
            <pubDate>Thu, 31 Jan 2019 03:35:04 GMT</pubDate>
            <atom:updated>2019-01-31T03:35:04.390Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>Big data — common pitfalls</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LW3cxR_wJnparcqK6UOcJQ.jpeg" /></figure><p>Recently, Volvo created an application called roam delivery service which allows deliveries to be made to your car while the car sits unattended in the car park. Couriers access your car to deliver the goods using a single use digital key, which is disabled once the delivery is complete. What this demonstrates is something we already know: the appetite for advanced information services is huge and expanding exponentially. These services are only possible because we’re able to leverage insights derived from massive volumes of data, generated both by people and ‘things’.</p><p>Did you know that we, the human race, now create a quintillion bytes worth of data every day? It’s impossible to even fathom what a quintillion of anything looks like; nothing in our everyday experience could possibly put these numbers into perspective. The name we use for all this information is ‘Big Data’ and according to the websites of many major advisory firms, it’s the most important new frontier for business.</p><p>There’s no question that organisations are now capturing unprecedented amounts of data. For this we can thank more effective data management systems; increased connectivity through smart phones and other devices; advances in satellite information and imagery; the ever increasing adoption of social networking, and; new technologies designed specifically to track customer’s personal information and preferences.</p><p>The usefulness of data is, however, not measured by its volume. Large, unfocused data collection exercises often yield relatively little new information. Recently we completed a major piece of work developing a new business model for a key Victorian State Government department. Before developing our model we tested the significance of over 250 variables which were believed to be important drivers of demand for the authority’s services. What we found was that only a handful provided any real value or significance from a forecasting perspective.</p><p>The current hype surrounding Big Data is almost entirely focused on techniques for collecting and storing data. Questions raised usually take the form: “Do I continue expanding the size of my local servers or do I leverage the cloud?” and “If I use the cloud, should I use the public cloud, or invest in a private cloud infrastructure or some hybrid?” This focus on technologies for collecting and storing large volumes of information is misplaced.</p><p>Collecting and storing information is important, but only to the extent that it is efficiently and effectively incorporated into everyday decision-making. Therein lies the problem for most organisations struggling to understand what to do about Big Data. Unless the right questions are asked, critical investments in data acquisition and management systems may be wasted.</p><p>The fundamental question we must ask ourselves is: “How do I use this data to make better decisions which will increase customer loyalty, improve productivity and save money?”</p><p>Basically the value of Big Data is determined by data analysis. Data analysis is what we do in order to provide value-added services, for example fridges and coffee machines that can keep inventory and reorder stock when running low. Predictive analytics take this one step further, using insights from Big Data to provide expectations of future outcomes based on any number of possible scenarios, such as the effect of a new competitor in the market or a shortfall in supply or demand for a product or service.</p><p>Despite the complex mechanics that sometimes underpin the analysis, all businesses really need to know is that, if used in the right way, Big Data (coupled with advanced data analysis) can result in increased profits and significant financial savings.</p><p>For organisations who want to capitalise on these insights (and any leader who doesn’t is going to be left behind) you need to grasp a few realities first:</p><ul><li>Most of the information being stored has negligible value</li><li>The only way to determine what is meaningful is through sophisticated data analysis</li><li>Off the shelf software packages and generic front-end Business Intelligence (BI) interfaces, which focus largely on data visualization, will not provide you with a competitive advantage (your competitors are using exactly the same tools)</li><li>Only predictive analytics (complex, adaptive statistical models) specifically tailored for your organization and operating environment will give you the full range of insights you need to prepare for every eventuality</li></ul><p>Far too often organisations, interested in harnessing the potential that predictive analytics can provide, focus on IT specifications. They spend millions on tools that are only superficially customised and do little to assist in decision making and planning. Insufficient emphasis is placed on the objective itself, the provision of useful predictions. Instead, these insights are often lost in the focus on visual functionality and systems integration. In addition, stock-standard statistical models are applied, which cannot adapt to valuable new sources of information derived from the gigantic Big Data reservoir, without significant further investment and adaptation.</p><p>Any analyst who assumes that insights into the year ahead can simply be derived from last year’s data will be disappointed. Millions of observations, including consumer sentiment, economic commentary, futures trading and speculation, as well as transaction data (such as ‘point of sales’ info) must be taken into account in order to develop truly useful forecasting models and simulators.</p><p>Unfortunately, the latest mathematical and statistical innovations can take years before they filter through to the commercial world and into easy-to-use off-the-shelf software packages. This isn’t good enough. Today’s commercial leaders need to seek out and take advantage of the latest statistical innovations as quickly as they can to stay ahead.</p><p>This is where Predictive Analytics Group can help. Our comprehensive analysis of what data is actually important, coupled with the most advanced adaptive statistical models on the market, provides us with a distinct advantage over those cookie-cutter data management and visualization tools.</p><p>The first thing leaders and managers need to ask themselves, in order to really harness the power of Big Data, is “How will these tools help me make better decisions?”</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1416d8fa8668" width="1" height="1" alt=""><hr><p><a href="https://medium.com/predictive-analytics-group/big-data-common-pitfalls-1416d8fa8668">Big data — common pitfalls</a> was originally published in <a href="https://medium.com/predictive-analytics-group">Predictive Analytics Group</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Profit Warning for Christmas Retailers]]></title>
            <link>https://medium.com/predictive-analytics-group/profit-warning-for-christmas-retailers-badc6fe2be96?source=rss-e016f0cb1b4e------2</link>
            <guid isPermaLink="false">https://medium.com/p/badc6fe2be96</guid>
            <category><![CDATA[big-data]]></category>
            <dc:creator><![CDATA[Predictive Analytics Group]]></dc:creator>
            <pubDate>Thu, 31 Jan 2019 03:30:46 GMT</pubDate>
            <atom:updated>2019-01-31T03:30:46.362Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0Fgt2OiX5DqDylZsOJrrIg.jpeg" /></figure><p>Retailers have been warned that unless they learn to properly analyse customer data, they run the risk of massively reduced profits during the Christmas period.</p><p>Big data analysts — Predictive Analytics Group say major retailers could be increasing sales by 100s of millions of dollars over the next six weeks, if they adapt available cutting edge data analysis techniques.</p><p>Predictive Analytics Group has combined advancements in mathematics (developed in-house) with state of the art technology to create a real time platform that cuts through the complexity of big data providing unique and powerful insights into the changing nature of the retail market and consumer behaviour.</p><p>“We have constructed a real-time monitoring app that allows the marketing manager in a company to walk up and down their shop aisle and see — with thanks to advanced data analytics — how demand for each product will change and what the price of each product should be,” Dr. Gazos said.</p><p>“Through loyalty card programs such as Fly Buys and Myer One as well as point of sales transactions data — retailers obtain huge amounts of information from their customers. But they are not unlocking from this data the full potential of what their customers are really telling them about what they want,” Dr. Gazos added.</p><p>Dr. Gazos went onto explain that big retailers already have large teams that analyse data, but says that in a lot of cases they are using out dated techniques that mask rather than inform how consumer preferences are changing.</p><p>According to Predictive Analytics Group by analysing the data in the right way, retailers can make the Christmas shopping experience more enjoyable for their customers.</p><p>“We know the stress Christmas shopping brings, if retailers learn to analyse data in the correct way, and respond more efficiently by pre-empting changes in consumer wants and needs, this extra stress can be minimised,” Dr Gazos said..</p><p>“Consumer preferences are changing, with the advancement in online shopping retailers need to adapt to the way they present their products and service more generally,” Dr Gazos added.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=badc6fe2be96" width="1" height="1" alt=""><hr><p><a href="https://medium.com/predictive-analytics-group/profit-warning-for-christmas-retailers-badc6fe2be96">Profit Warning for Christmas Retailers</a> was originally published in <a href="https://medium.com/predictive-analytics-group">Predictive Analytics Group</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Using Analytics to Manage Risk]]></title>
            <link>https://medium.com/predictive-analytics-group/using-analytics-to-manage-risk-d496fc98d341?source=rss-e016f0cb1b4e------2</link>
            <guid isPermaLink="false">https://medium.com/p/d496fc98d341</guid>
            <category><![CDATA[big-data]]></category>
            <dc:creator><![CDATA[Predictive Analytics Group]]></dc:creator>
            <pubDate>Thu, 31 Jan 2019 03:25:46 GMT</pubDate>
            <atom:updated>2019-01-31T03:25:46.166Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5KGBbo4LeqC0E7t2eYZXYQ.jpeg" /></figure><p>As the commercial world changes at an unprecedented rate, the single biggest risk faced by many organisations is that of market failure through irrelevance.</p><p>That’s where so-called “uncertainty modelling” comes in — analysing data to create useful, workable visions of the future that enable your organisation to remain indispensable to your customers…</p><p>Increasing numbers of companies across the world are recognising that the only way to ensure their continuing relevance and survival is to actively build their organisations’ market strategy around satisfying the customer’s changing needs..</p><p>They can do this in many ways, from taking major strategic decisions such as developing new channels to market and revamping product line-ups to comparatively minor pricing and packaging changes.</p><p>What unifies all such decisions is the shared need for them to be based on the correct analysis of accurate and relevant customer data that provides an effective basis for a long-lasting and productive relationship.</p><p>That is because, in this era of unparalleled technological advancement and the increasingly empowered consumer, being able to identify and trace emerging trends and new preferences has never been more important.</p><p>Fortunately, the ability to see into the future is moving out of the world of “guestimations” and into the sphere of increasingly precise analysis, based on custom built software applications that give companies real-time feedback on key metrics and help them mitigate the risk of making the wrong decisions.</p><p>Analytics that can reduce uncertainty through accurate modelling have the potential to become an enormously valuable company asset, which drives new product development and strategic marketing initiatives.</p><p>They also enable companies to identify obsolescence, to predict the impact of price changes and promotional initiatives, and to estimate the likely impact of new information.</p><p>In other words, while nothing that remains in the future can ever be predicted with 100% certainty, such analysis tools can do much to eradicate the risk of alienating future consumers, either through inaction or taking the wrong action..</p><p>Companies that are serious about ensuring their long-term relevance by building better customer relationships need to get to grips with this source of potential advantage before their competitors do.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d496fc98d341" width="1" height="1" alt=""><hr><p><a href="https://medium.com/predictive-analytics-group/using-analytics-to-manage-risk-d496fc98d341">Using Analytics to Manage Risk</a> was originally published in <a href="https://medium.com/predictive-analytics-group">Predictive Analytics Group</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Big Data — Why all the Hype?]]></title>
            <link>https://medium.com/predictive-analytics-group/big-data-why-all-the-hype-f65d2fab3f88?source=rss-e016f0cb1b4e------2</link>
            <guid isPermaLink="false">https://medium.com/p/f65d2fab3f88</guid>
            <category><![CDATA[predictive-analytics]]></category>
            <category><![CDATA[big-data]]></category>
            <dc:creator><![CDATA[Predictive Analytics Group]]></dc:creator>
            <pubDate>Thu, 31 Jan 2019 03:21:03 GMT</pubDate>
            <atom:updated>2019-01-31T03:21:03.955Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>dBig Data — Why all the Hype?</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eWYm3Dm7s0t8lLWYRW4pow.jpeg" /></figure><p>According to Gartner, Big Data is: “is high-volume, high-velocity and high-variety information assets that demand cost-effective, innovative forms of information processing for enhanced insight and decision making”.</p><p>Which says it all really — there’s lots of it, it’s complicated, and it can help you make the right decisions provided you use the right means of analysis.</p><p>Because somewhere in your mound of data are the key facts whose interrelations can help you to build meaningful forecasts of market trends and consumer behaviours.</p><p>The problem lies in finding the right way of analysing it — while there are plenty of software packages that help you to store and centralise massive amounts of data from a vast range of sources, being able to do something useful with it is still the key challenge.</p><p>And using the wrong solution can all too easily produce the wrong results. For instance, using standard BI software is never going to be appropriate — it’s too elementary and doesn’t take into account the data’s all-important underlying properties which are crucial in the generation of forecasts.</p><p>But there is an emerging answer, which many are backing as favourite to become the go-to Big Data solution. This is Predictive Analytics, a critical combination of statistical, quantitative, IT and market expertise.</p><p>Right now, there is an early adopter advantage open to those companies that take the opportunity to gain greater insight into their Big Data assets by using Predictive Analytics.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f65d2fab3f88" width="1" height="1" alt=""><hr><p><a href="https://medium.com/predictive-analytics-group/big-data-why-all-the-hype-f65d2fab3f88">Big Data — Why all the Hype?</a> was originally published in <a href="https://medium.com/predictive-analytics-group">Predictive Analytics Group</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Announcing PAG’s Retail Toolkit]]></title>
            <link>https://medium.com/predictive-analytics-group/announcing-pags-retail-toolkit-450933f55dc6?source=rss-e016f0cb1b4e------2</link>
            <guid isPermaLink="false">https://medium.com/p/450933f55dc6</guid>
            <category><![CDATA[retail]]></category>
            <category><![CDATA[predictive-analytics]]></category>
            <category><![CDATA[big-data]]></category>
            <category><![CDATA[analysis]]></category>
            <dc:creator><![CDATA[Predictive Analytics Group]]></dc:creator>
            <pubDate>Thu, 31 Jan 2019 03:12:46 GMT</pubDate>
            <atom:updated>2019-01-31T03:12:46.562Z</atom:updated>
            <content:encoded><![CDATA[<h4><strong>Optimising pricing strategies for retailers</strong></h4><p>How can big data give your business a competitive edge:</p><p><strong>• Product pricing optimisation</strong></p><p>Determine optimal price levels for each product at a given time.</p><p><strong>• Determine the most effective promotional strategies</strong></p><p>Empirical evidence for determining optimal promotion campaigns to achieve strategic goals.</p><p><strong>• Smart visual merchandising</strong></p><p>Bespoke algorithms utilising affinity analysis and customer preference information to maximise opportunities at each point-of-sale.</p><h4><strong>The big data advantage strategies for retail</strong></h4><p>The success of retail hinges on a multitude of competing factors which can be difficult to map</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Y0rhsKaBe8rYlqoA2EixJA.png" /></figure><p><strong>The quantity sold of a given product should be a function of:</strong></p><ol><li>Product prices (including price of complements and substitutes)</li><li>TV, advertising campaigns</li><li>Consumer behaviour by geographic region</li><li>Seasonality and major events i.e. Christmas, Easter</li></ol><h4><strong>Product pricing optimiser:</strong></h4><p>PAG’s Retail Toolkit utilises algorithms that can reveal relationships across thousands of items. It can</p><ul><li>calculate the best time to apply a price markdown</li><li>Models complements and substitutes and take into account the impact of cross-price effects when determining product pricing</li><li>Account for clusters and non-linearity. This is crucial for sensibly assessing the impact of product prices.</li></ul><h4><strong>Model Promotional Strategies</strong></h4><p>PAG’s Retail Toolkit allows retailers to measure the effectiveness of promotional strategies with confidence</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/616/1*Cg9MLKe8gj_h7oXu0cVCcw.png" /></figure><p><strong>Problem: the effectiveness of promotions is often hard to predict</strong></p><p>Without proper robust modelling (taking into account costs, changing consumer preferences, and economic conditions), promotional strategies can often be extremely ineffective and costly</p><p><strong>Solution: Manage uncertainty and prioritise best available strategy</strong></p><p>Consider the following example:</p><p>Retailer X’s objective is achieving at least $2 million in sales over the next two quarters PAG’s Retail Toolkit considers two promotional strategies:</p><p>1. Strategy A is expected to return $5 million with a lower bound of $1.5 million<br>2. Strategy B is expected to return $4 million with a lower bound of $2.5 million</p><p>If the objective is achieving at least $2 million in sales then Strategy B is preferred.</p><h4><strong>Smart Online Visual Merchandising</strong></h4><p><strong>PAG’s Retail Toolkit allows online retailers to maximise the opportunities at each point-of-sale:</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/362/1*7rHrNlyDHkfdzfJuZcssiQ.png" /></figure><p><strong>Problem: Generic product placement is not effective</strong></p><p>Product placement techniques such as ‘we recommend’ or ‘other customers also bought’ can be too generic</p><p><strong>Solution: Each product web page is it’s own unique set of cross-selling and up-selling opportunities.</strong></p><p>PAG’s underlying algorithm utilises affinity analysis which models complements and substitutes and incorporates individual consumer preferences to create a unique ‘store’ for each individual consumer</p><p>Each product page a consumer enters will have unique product placement, based on the following variables:</p><ul><li>price of other goods (cross-price elasticity)</li><li>existing promotions</li><li>overall consumer purchase behaviour (‘association rule learning’ e.g. people who buy product X tend to buy product Y)</li><li>individual consumer browsing and purchase data (where available)</li></ul><p>By analysing these variables together, retailers are able to maximise the opportunities at each point-of-sale</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=450933f55dc6" width="1" height="1" alt=""><hr><p><a href="https://medium.com/predictive-analytics-group/announcing-pags-retail-toolkit-450933f55dc6">Announcing PAG’s Retail Toolkit</a> was originally published in <a href="https://medium.com/predictive-analytics-group">Predictive Analytics Group</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>