<?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 Dany Srage on Medium]]></title>
        <description><![CDATA[Stories by Dany Srage on Medium]]></description>
        <link>https://medium.com/@dany.srage?source=rss-aec0d818da1f------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*PrU5ElXP0DhKjHja3YkzHw.jpeg</url>
            <title>Stories by Dany Srage on Medium</title>
            <link>https://medium.com/@dany.srage?source=rss-aec0d818da1f------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 08:01:51 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@dany.srage/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[Explainability of the features? No! Of the hyperparameters.]]></title>
            <link>https://medium.com/defactoblog/explainability-of-the-features-no-of-the-hyperparameters-ad797918155f?source=rss-aec0d818da1f------2</link>
            <guid isPermaLink="false">https://medium.com/p/ad797918155f</guid>
            <category><![CDATA[hyperparameter-tuning]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[explainable-ai]]></category>
            <dc:creator><![CDATA[Dany Srage]]></dc:creator>
            <pubDate>Fri, 15 Mar 2024 10:48:35 GMT</pubDate>
            <atom:updated>2024-03-15T15:31:57.833Z</atom:updated>
            <content:encoded><![CDATA[<p>As data scientists, we enjoy training models, but the enthusiasm often fades when it’s time to fine-tune the model’s hyperparameters after preparing all the features.</p><p>We’ve all been in situations where a project stakeholder asks if we can add a feature or change tiny details but it turns out it’s much more time-consuming than they thought because we need to run the optimization again.</p><p>Although there’s an abundance of definitions for the hyperparameters of various models, translating them into practical implications can be challenging. <strong>The purpose of this article is to provide a tangible method and understanding of how hyperparameters influence model performance, highlighting the crucial ones and explaining their interactions</strong>. It also recaps how to explain an algorithm with respect to its features.</p><p>Even though on their own each hyperparameter is very important, since many of them contribute to reducing overfitting versus model complexity, we need to find a balance between all of them and I hope this analysis will help in finding it.</p><h3>M<strong>ethodolog</strong>y</h3><p>Models that we put in production need to be explainable: we need to understand how each feature impacts the overall predictions.</p><p><strong>The whole idea of this article is that instead of explaining the features, why wouldn’t we explain the hyperparameters?</strong> That way, we could see how each one impacts the model performances.</p><p>Indeed, when we take models such as linear regressions, the prediction function looks like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/458/1*9r9ZZLf6ElyBqXCIbrr8zg.png" /></figure><p>where the w_i are the coefficients and x_i are the features. If you think about it,<strong> the features and the parameters of the models are almost symmetrical.</strong> So what if we swap their roles in our explanation models?</p><p>A common method to interpret a model is to use the SHAP approach. It’s based on game theory and in short, for each prediction, it tells how each feature contributed to it. You can find more <a href="https://shap.readthedocs.io/en/latest/">information here</a>.</p><p>To use this method, <strong>we first need to build a Machine Learning model that predicts the model performances from the hyperparameters</strong>. Once we validated that this model gave very good predictions, we can observe and learn from its SHAP summary. We’ll see below an example.</p><p>We’ll focus on <a href="https://xgboost.readthedocs.io/en/stable/#">XGBoost</a> algorithm as it’s one of the most widely used algorithms and has many hyperparameters. To illustrate our findings, we conducted analyses on both an internal dataset and the renowned <a href="https://www.kaggle.com/c/titanic/data">Titanic dataset</a>. While a comprehensive analysis across multiple datasets would be ideal, we hope this analysis will still help better understand these hyperparameters and the methodology behind. Being classification problems and our internal dataset being imbalanced, I focused on the average precision metric.</p><p>To keep the article concise, we won’t delve extensively into the definitions of all the parameters of XGBoost, as you can find detailed descriptions in the official documentation <a href="https://xgboost.readthedocs.io/en/stable/parameter.html">here</a>.</p><h3>Let’s now look at the SHAP summaries:</h3><p>Generating these summaries is as easy as the following in Python.</p><pre>import shap<br><br>explainer = shap.TreeExplainer(model) # model is a trained sklearn model where <br>				      # y_i are the performances of the model<br>				      # and x_i are the hyperparameters<br>shap_values = explainer.shap_values(X_test) # X_test is our test data<br>shap.summary_plot(shap_values, X_test)</pre><p>Below you can see the SHAP summaries for the two datasets we’ll analyze.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/781/1*2WaEkQVhDBPXO3ul5JOJog.png" /><figcaption><strong>SHAP Summary for the Titanic dataset</strong></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/718/1*NF_zUT0bafOarJkL4bykUg.png" /><figcaption><strong>SHAP Summary for our internal dataset</strong></figcaption></figure><p><strong>How to read SHAP summaries?</strong></p><p>As a reminder of how SHAP summaries work, each line on this graph represents a feature, and features are sorted by importance. Each dot for a given feature represents a single prediction and its position on the x-axis represents its marginal impact on the prediction. If a feature has a SHAP value of -0.01, it means that this feature value causes the prediction to decrease by this amount.</p><p>On the right, you can find a scale for feature values. If the dot is red, it means that the value for this feature is very high. And if this point is more to the right, it means that high values for a feature impact positively the model.</p><p>Note that when many predictions have the same contribution for a given feature, that’s when the line starts to get much larger (the points start agglomerating on the summary).</p><p><strong>Insights and learnings</strong></p><p>Before diving into each hyperparameter one by one, the first insight we have is that the most important features to optimize are:</p><ul><li><strong>learning rate, min_child_weight, colsample_by_tree and the max depth</strong></li></ul><p>while the least important one seems to be:</p><ul><li>gamma</li></ul><p>💡Note how the impact of <em>min_child_weight and n_estimators</em> differs from one dataset to the other.</p><p>Let’s now see what we can learn about each parameter one by one. For this, we generated SHAP dependency plots using the following:</p><pre>for i in range(len(X_test.columns)):<br>    shap.dependence_plot(i, shap_values, X_test)</pre><h3>Insights about hyperparameters:</h3><p><strong>Learning rate</strong>: it seems to be both times the most important feature. The red part of the graph is on the left, meaning that the larger the learning rate is, the less performing the model.</p><p>This parameter controls how each new tree in the boosting will learn from the residual error. A smaller value means that the model will take more time to learn and will need more trees in the boosting. A higher value on the opposite will converge faster but might miss an optimal result. It’s therefore why <strong>according to our Shapley values, the feature it interacts the most with is the number of estimators (see the n_estimators part and graph).</strong></p><p>We can see below the dependency plot of the performances as a function of the learning rate. We can see that the highest learning rate(purple line) tends to overfit quickly, while <strong>the most performing models seem to be the one with the lowest learning rates, even though it requires more iteration to get the best performances</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/765/1*O5HXP70fsjGEoUf0j9q9KA.png" /></figure><p><strong>Max depth: </strong>it’s the maximum depth of each tree. As can be seen in the plot below, a value too high will lead to overfitting because the tree will be too complex. On the opposite, if too low, the model will be too simple leading to worse results.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/592/1*MnOp9lH0USvGVe0s_J1XZA.png" /></figure><p><strong>colsample_by_tree</strong>: the ratio of features to sample for each new tree. SHAP automatically displays the features it interacts the most with by coloring each point with this feature value (see graph below). For this feature, we can see in the figure below that the feature it interacts the most with is the learning rate. It seems that if the number of columns to sample is low (0.5), having a lower learning rate will help the model to perform better.</p><p>Intuitively, when the column sampling is low, each tree is very different from each other. It is, therefore, better and “more stable” to learn slightly from each additional tree, hence a lower learning rate being optimal.</p><p>This shows that if one wants <strong>to have a low sampling rate, it’s important to keep lower learning rates in our search space.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HKOfGVyDvIP_-fV-s4p7YQ.png" /></figure><p><strong>n_estimators: </strong>the number of trees in our boosting algorithm. There are two learnings: the more trees, usually the better. One could easily <strong>cut the space to search for hyperparameters by stopping when the below curve becomes flat:</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/585/1*_bhHKP4lTL_cJ1memUJElw.png" /></figure><p>The second learning is that the parameter it interacts with the most is the learning rate as discussed above. <strong>With a low number of estimators, it’s better to have a high learning rate because the model won’t have too much time to learn during training</strong>. On the contrary, with a high number of estimators, we can afford a lower learning rate which gives better performances overall.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mCm20nvte2s84bHso3XAhQ.png" /></figure><p><strong>subsample</strong>: the ratio of training data to sample before training the tree. We see very different behaviors between the two datasets, where in one the higher the better, and in the other the reverse is true. <strong>Subsample seems to highly depend on the dataset</strong>. The parameter it interacts the most with is the learning rate. The reason is similar to colsample_by_tree.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xvy22-jxL3w4N0mG1D-vGw.png" /></figure><p><strong>min_child_weight</strong>: controls how much “instance weight” is needed in a child, with a higher value meaning less split and thus less overfitting. It seems much more important in the Titanic dataset than the other one.</p><p>Respectively we can see its impact on our internal dataset and the Titanic dataset in the graphs below:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/675/1*CD_1BylAbcrm8-Tr25VzIg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/674/1*Q40Rk9E2K1nQo6iTPTGSfQ.png" /></figure><p>Two things strike me:</p><ul><li>the feature with which it interacts the most changes with the dataset.</li><li>High values always decrease performances on the Titanic dataset while this is not true in our internal one.</li></ul><p>This parameter seems thus highly dependent on the dataset.</p><p><strong>gamma</strong>: it represents the minimum loss reduction needed to split the tree. In both datasets, this seems to be the less impactful parameter and <strong>for our use case, it could be dropped to help reduce the search space when trying to get the optimal solution.</strong></p><h3>Validating the results</h3><p>Since this analysis is based on a model that itself has some flaws, I wanted a more classic method to understand how each hyperparameter impacts the model performance. For each hyperparameter, I created a partial dependency plot with respect to every other studied hyperparameter and I then looked at the range of the evaluation metric. The hypothesis is that if a hyperparameter is important, changing its value leads to an important variation in the evaluation metrics. We got the following results:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pzI0uyhb_xzikKe0RoIZHA.png" /></figure><p>Results were very similar except for the Titanic dataset where min_child_weight appears to be much more important than in SHAP.</p><h3>Conclusion</h3><p>I hope that with this article you get:</p><ul><li>a better understanding of how to study the explainability of models.</li><li>a better understanding of each hyperparameter and how changing their values can impact the model results.</li><li>an overall method to get a global vision of your search space of hyperparameters and how to optimize it.</li></ul><p>As a recap, my top learnings from this analysis are:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EIZLEpvj7k0ACRS8lpBsSQ.png" /></figure><p>Obviously, these learnings might differ with different datasets, and running such an analysis with your dataset might be helpful to save time down the line.</p><p>There are other methods to optimize hyperparameters that were not covered in this article such as Bayesian Search which are also worth looking into.</p><p><em>Thanks to Yohan, Data Scientist at Defacto, and Marc-Henri, CTO of Defacto, for their feedback.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ad797918155f" width="1" height="1" alt=""><hr><p><a href="https://medium.com/defactoblog/explainability-of-the-features-no-of-the-hyperparameters-ad797918155f">Explainability of the features? No! Of the hyperparameters.</a> was originally published in <a href="https://medium.com/defactoblog">get-defacto</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How BlaBlaCar uses Machine Learning to maximise the success of new drivers]]></title>
            <link>https://medium.com/blablacar/how-blablacar-uses-machine-learning-to-maximise-the-success-of-new-drivers-4b2ca1e7e7c6?source=rss-aec0d818da1f------2</link>
            <guid isPermaLink="false">https://medium.com/p/4b2ca1e7e7c6</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[engineering]]></category>
            <category><![CDATA[tech]]></category>
            <category><![CDATA[crm]]></category>
            <category><![CDATA[modeling]]></category>
            <dc:creator><![CDATA[Dany Srage]]></dc:creator>
            <pubDate>Wed, 24 Mar 2021 15:02:04 GMT</pubDate>
            <atom:updated>2021-03-24T15:02:04.541Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VXlUslZP811G4hTeHmHr4g.jpeg" /></figure><p><em>Article co-written with </em><a href="https://medium.com/@antoine.barczewski_830"><em>Antoine Barczewski</em></a><em>, Data Scientist at BlaBlaCar.</em></p><p>When a member publishes their first trip, our goal is to help them make the most out of the platform by maximizing their chances of getting passengers. We realized that we could achieve this goal by using data to create personalized recommendations and tips. Our recommendations are then pushed by CRM, whose role is to increase member engagement through education and conversion all along the member lifecycle.</p><h3>The landscape of explainable models …</h3><p>To provide these recommendations to new drivers, we decided to build an explainable model which predicts success. That way, we can understand which features are driving a higher success rate and make recommendations based on these. Our assumption that is being A/B tested is that spotting one or two important features will lead to personalized recommendations and that this would have more engagement than generic guidelines.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/491/0*fwYKgxlABxU1YKNb" /><figcaption>Example of recommendations our algorithm would produce to the end user.</figcaption></figure><p>Looking at the landscape of explainable AI, we identified various methods which can explain and quantify how each feature relates to the predicted outcome. Here are examples:</p><ul><li>Building a model which is interpretable out-of-the-box such as linear models or decision trees. That way, we can easily read on the model itself how each feature impacts the success. We did not use this method because we wanted to use more complex models.</li><li>For more complex models, the LIME<a href="https://arxiv.org/abs/1602.04938">[1]</a> ( Local Interpretable Model-Agnostic Explanations) method states that we can approximate any black-box model to a linear model locally. Hence, for each training sample, we could use this linear approximation to interpret how moving different parameters impact success.</li><li>Also for more complex models such as gradient boosted trees, a method called SHAP developed by Lundberg and Lee <a href="https://arxiv.org/abs/1705.07874">[2]</a> relies on what is called SHAP Values from game theory. The goal of SHAP values is to find how much each feature contributed to our success probability (more info <a href="https://christophm.github.io/interpretable-ml-book/shapley.html">here [3]</a>). For instance, if we predict that a trip’s success is 80% while our average success rate is 50%, each feature has a SHAP value which explains part of this 30% gap. We can then use these numerical values to analyze our model results and see if one or two features are significantly driving up/down the success score.</li><li>Another method that we thought would be interesting would be to use the partial dependence plot of each feature <a href="https://scikit-learn.org/stable/modules/partial_dependence.html">[4]</a>. The idea of these plots is to see the marginal effect of a single variable on the success variable. This assumes that every other feature is independent of the one we look at. This helps see how moving this feature moves on average the success rate. Then, for a single point, we would see where we stand on our partial dependence plot to make a prediction. However, these plots are based on aggregated data. It can therefore only give general recommendations.</li><li>The method above is not helpful in giving individualised recommendations. However, in the same spirit, for every data point, we can try to move all features within reasonable bounds, predict success again and see how moving them impacts the score. The one with the higher impact will be the recommended feature. This also assumes that our features are independent of each other. This method is similar to a tool called What-if [5] Tool.</li></ul><p>We decided to use this last method to produce individual recommendations as we thought it replicated well what users would do in real life and it was easy to interpret. However, we also produced SHAP values regularly to monitor our model behaviour, to create insights, and to spot inconsistency in our data.</p><h3>… And how we chose the right one</h3><p>So let’s recap what we did from the start. First, we built a model to predict drivers’ chances of success. This model for each publication is based on our knowledge, experience, and data. To build it, we’ve trained a gradient boosting algorithm on 2019 data. This dataset gathers features from the drivers’ profiles and the trips’ descriptions. The target tells whether or not the publication experienced success. Gradient boosting is a well-known machine learning algorithm that showed good performance in our case. The area under the curve was around 0.8. Looking at the precision-recall curves, we validated that the model was a good proxy of success.</p><p>To build the model with the “what if” methodology in mind, we had to identify three groups of features:</p><ol><li>features that help prediction performance,</li><li>actionable features at publication time,</li><li>actionable features at departure time.</li></ol><p>Indeed, if we provided only features which are actionable to the user, we did not have the best model. We, therefore, needed to have these features that help the model perform better. We also needed actionable features that could be potential recommendations. There are also two types of such actionable features. The ones actionable right after the publication, and the ones that will help a future publication. For instance, sending an email saying that you should have published this trip two days earlier would not help this time. However, this information can be shared after departure to help for the next publication.</p><p>We noticed two issues. When there is collinearity between features, one feature can hide the true effect of another feature because the model has learned it this way. For instance, we saw that the number of days between signup and publication had a huge impact on success based on our model. However, the feature telling whether or not a member experienced success as a passenger before publication turned out to have very little importance. Looking at the various SHAP values (main effect and interactive values) of these features, we understood the interactions that these features had. Indeed, this “number of days between signup and publication” feature is not actually useful unless it hides that the new driver was already a passenger, with therefore a few ratings boosting their chance of success as drivers. We addressed the issue by grouping features like these two in groups and summing up their SHAP values to understand the impact of this group on success. Because of the additivity of the SHAP values, we knew the impact of these features as a group and did not have to worry about the interactions of these grouped features between each other.</p><p>The second problem coming from collinearity is that when we are changing a feature A, any feature that depends on A because of collinearity remains unchanged in our simulation. However, in real life, a member changing feature A would also be changing other collinear features. Hence, it would be changing the success probability differently from our simulation. It was not a problem for us because our actionable features were independent of every other feature, actionable or not.</p><h3>Conclusion</h3><p>At the end of the day, we’ve built two separate models, a sort of “what if tool” that pushes recommendations of actions to newbie drivers at publication and departure time and a SHAP-based model that produces aggregated SHAP values to support CRM analysis.</p><p>These two models are built upon one simple classification model. We believe that the true value doesn’t lie in the prediction model here but more in its explanation. Understanding a model behaviour is something we always work on at BlaBlaCar. Here, it enabled us to help the members take the most advantage of the platform and maximize their chances of success.</p><p>Thanks a lot to Akshita, Célia, Emil and Guendalina for their very important contribution to the project as well as Emmanuel, Nicolas and <a href="https://medium.com/@emilie.baliozian">Emilie</a> for their help in writing the article and Alizée for the illustration.</p><h3>References:</h3><p>[1] Marco Tulio Ribeiro and Sameer Singh and Carlos Guestrin. 2016. ““Why Should I Trust You?”: Explaining the Predictions of Any Classifier”. <a href="https://arxiv.org/abs/1602.04938"><br>arXiv:1602.04938</a> [cs.LG] <a href="https://arxiv.org/abs/1602.04938">https://arxiv.org/abs/1602.04938</a></p><p>[2] Scott Lundberg and Su-In Lee. 2017. “A Unified Approach to Interpreting Model Predictions”. <a href="https://arxiv.org/abs/1705.07874">arXiv:1705.07874</a> [cs.AI] <a href="https://arxiv.org/abs/1705.07874">https://arxiv.org/abs/1705.07874</a></p><p>[3] Molnar, Christoph. “Interpretable machine learning. A Guide for Making Black Box Models Explainable”, 2019. <a href="https://christophm.github.io/interpretable-ml-book/">https://christophm.github.io/interpretable-ml-book/</a>.</p><p>[4] Scikit-learn: Machine Learning in Python, Pedregosa et al., JMLR 12, pp. 2825–2830, 2011. <a href="https://scikit-learn.org/stable/modules/partial_dependence.html">https://scikit-learn.org/stable/modules/partial_dependence.html</a></p><p>[5] <a href="https://pair-code.github.io/what-if-tool/">https://pair-code.github.io/what-if-tool/</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4b2ca1e7e7c6" width="1" height="1" alt=""><hr><p><a href="https://medium.com/blablacar/how-blablacar-uses-machine-learning-to-maximise-the-success-of-new-drivers-4b2ca1e7e7c6">How BlaBlaCar uses Machine Learning to maximise the success of new drivers</a> was originally published in <a href="https://medium.com/blablacar">BlaBlaCar</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Meet Dany, who joined BlaBlaCar as a Data Scientist during lockdown]]></title>
            <link>https://medium.com/blablacar/meet-dany-who-joined-blablacar-as-a-data-scientist-during-lockdown-ff34fef0f5cc?source=rss-aec0d818da1f------2</link>
            <guid isPermaLink="false">https://medium.com/p/ff34fef0f5cc</guid>
            <category><![CDATA[culture]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[tech]]></category>
            <category><![CDATA[onboarding]]></category>
            <category><![CDATA[data]]></category>
            <dc:creator><![CDATA[Dany Srage]]></dc:creator>
            <pubDate>Wed, 28 Oct 2020 08:29:12 GMT</pubDate>
            <atom:updated>2020-10-28T08:29:12.038Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>Meet Dany, who joined BlaBlaCar as a Data Scientist during the Covid-19 lockdown</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*kBKtwBXKIo3nHqms" /></figure><p><strong>Hi Dany. Tell us about yourself.</strong></p><p>I’m Dany, and I’m currently a Data Scientist at BlaBlaCar. Our team’s mission is to improve the efficiency of other teams by automating data flows and decision making. This includes, for instance, detecting fraud, assisting marketing decisions, or designing new bus networks for BlaBlaBus.</p><p>Before joining BlaBlaCar, I studied engineering in France and the United States. After graduating, I joined Amazon as a Software Engineer. My first job was first to help integrate more Alexa in a smart home system (try it out!) and then I worked on the search engine. Missing <em>croissants</em> too much, I decided to go back to France — and what better timing than a pandemic to change a job and country?</p><p><strong>Why BlaBlaCar?</strong></p><p>I wanted to work in a tech-first and international environment. Working at Amazon, I realised that I liked working on products that I enjoy using since I can relate to users. BlaBlaCar checked these boxes. If you add the positive environmental impact and the Fun &amp; Serious culture of BlaBlaCar, I was convinced it was a good fit.</p><p>My favorite thing so far, besides adding “BlaBla” as a prefix to a lot of nouns, is the room for human interaction and collaboration. There are many opportunities to work with people across departments, and what’s particularly exciting is how closely I get to work with the Product team. In that sense, you get to really understand how your work fits in the overall product and how it can better serve our community of travelers.</p><p><strong>How was your remote integration experience?</strong></p><p>The good news about joining a team during lockdown was that there would be no “first day at work anxiety” since there was literally no first day <em>at </em>work and no need to worry about which shoes to wear. The not-so-good news is that at that time I thought it would be hard to integrate with the team remotely. But I had multiple onboarding sessions to get me on track:</p><p><strong>#1 Team onboarding</strong><br>It consisted of 3 different parts:</p><ul><li><em>Project:</em> I discovered and started working on my first project and had many interactions with the person handing me the work. This was the opportunity to understand the stack we are using, such as Google Cloud Platform’s Cloud Run to run the models on, as well as other Data Workflows.</li><li><em>Team buddy:</em> I had a teammate to whom I can ask any question regarding the team, the company… literally everything that crossed my mind. He gave me all the tips to get started.</li><li><em>Meetings: </em>I had many meetings during my first weeks to learn about the different topics my teammates worked on and the tools they were using. This was for instance the opportunity to understand how our data warehouse works, which data is available and when, and how it is computed. I was able to hit the ground running with clear info on how to navigate our data and the ability to learn about tools I was using for the first time.</li></ul><p><strong>#2 Company Onboarding Week</strong><br>I learned about the ins and outs of the company: strategy, teams, culture, you name it. Due to the context, this one happened 2 months after I joined and was 100% remote, which means that almost every day for a few weeks, I had many video meetings where I would learn about our different services (BlaBlaSure, BlaBlaLines…) and everything BlaBla-related. One of the very cool parts was to embody my favorite value, <em>Be The Member. </em>With other newbies, we had a day off to pick a city near Paris and try out the product with all of the necessary sanitary precautions.</p><p><strong>#3 Engineering Onboarding</strong><br>This one tackled the technical aspects of different teams and their stack. As Data Scientists, the stack we use is different from other engineers’. It was therefore the opportunity to learn about processes (<a href="https://medium.com/blablacar-tech/the-more-devs-the-merrier-part-1-e3fd041c0a10">more here</a>) that other teams have in place and to understand the ownership of the different APIs.</p><p>Prior to this onboarding, the different services I heard about were mostly the ones related to data-related projects. It therefore helped me understand the big picture and not only limited to our use case. One of my favorite parts was to learn more about the Foundations teams, which deploy so many tools I didn’t know about to help engineers be more efficient and overall have a reliable service.</p><p><strong>What were some challenges you experienced during your integration and how did you overcome them?</strong></p><p><em>Integrating: </em>When you start working remotely, you don’t really get the chance to discuss with people outside of your team or scope of work. Having these onboarding sessions was a nice way to chat with other newbies, teams, and even our CTO @Olivier Bonnet. What helped the most was the morning coffee chats with my team to discuss anything but work for 15 minutes a day. This was the best opportunity to get to know people on the team rather than only their SQL queries. After a few weeks of these meetings, I felt I knew them in real life (but with more pixels).</p><p><em>Team culture: </em>remotely, it’s harder to detect small, but important team habits: how quickly can we expect to have an answer to our messages? How often do we ask questions, on Slack or in person? Thankfully, BlaBlaCar has a really strong culture of transparency, which allows us to ask these questions.</p><p><em>Asking questions: </em>as a newbie, it’s harder to ask the small questions that might not warrant a Slack message or would be faster in person. It makes it more challenging to get access to information. But it allows you to really think twice about your questions: many get answered digging one extra minute into the code, and I could do it at my pace.</p><p><em>Technical discussions: </em>online meetings were helpful to get to know how we compute many metrics and data, but we often missed the meeting’s room whiteboard to go over different modeling approaches.</p><p><em>Hardware: </em>Unfortunately I had to wait a few weeks before getting my work laptop and had to start working on mine. Usually, I first set up my dev environment, which is time-consuming. But since it was only a temporary environment, I had to sort between the things what could wait and what couldn’t, and get creative about what could be on the cloud, with pros and cons, instead of on my local machine. It was, for instance, the opportunity to try the AI Platform Notebooks on GCP to avoid working on my machine and share work more easily. Today, the situation is much better and the team is doing a great job. Newcomers have their hardware on Day 1: a good example of our value “Fail, Learn, Succeed”.</p><p><strong>What advice would you give to someone joining a company remotely?</strong></p><p><em>Attend meetings/talks/events: </em>What helped me get started remotely was to use any opportunity for social interaction. “Do you have one minute for a call?”, someone asks on Slack, “sure I do!” was always the answer. At BlaBlaCar we are lucky to have many demos or company-wide talks which is another opportunity to learn about the company.</p><p><em>Ask questions: </em>Everyone in the company is very friendly and willing to help. Many of these questions led to calls, which led to more questions, more learnings, and more social interactions.</p><p><em>Ask them in bulk: </em>To be mindful of everyone’s time, I decided to write down all the questions that I had over time, and ask them all at once to limit the disruption. I found that by waiting for all the non-critical ones, many got answered with time, which reduced their number from 10 to 5. Eventually, in discussing answers to some questions, it answered more questions. At the end of the day, I would only have 3 left.</p><p><em>Give feedback</em>: Between all the meetings and 1:1s, there is always the opportunity to say when something feels wrong, or, equally important, when something is good. Since onboarding remotely is new to a lot of people, it might not be perfect. Therefore, sharing feedback always helps. For instance, I mentioned to my manager that waiting a couple of months for our general onboarding to learn about the culture and the company processes was not ideal. He therefore specifically crafted an onboarding document to help me understand all this ahead of time, which was really helpful.</p><p>Overall, I am really glad I joined BlaBlaCar. The culture of the company and everyone in my team made it very easy for me to get started and helped me find solutions to all the challenges I faced onboarding remotely. By the way, did I mention that <a href="https://careers.smartrecruiters.com/BlaBlaCar">we’re hiring</a>?</p><p>Thanks to Emilie Baliozian who helped me a lot with this article and to everyone who provided feedback.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ff34fef0f5cc" width="1" height="1" alt=""><hr><p><a href="https://medium.com/blablacar/meet-dany-who-joined-blablacar-as-a-data-scientist-during-lockdown-ff34fef0f5cc">Meet Dany, who joined BlaBlaCar as a Data Scientist during lockdown</a> was originally published in <a href="https://medium.com/blablacar">BlaBlaCar</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>