What are the correlations between different Belgian consumer profiles and their spending habits? How can these be used to improve Delhaize's collect service?

The goal of the following report is a mix between finding commercially interesting results and seeing how we can apply the classical methods and algorithms of data science to the database of a famous retail store brand, Delhaize. The results that we obtain are focused around the customers of Delhaize in Belgium that use an online shopping service called the Collect service. The collect service allows customers to shop online and pick up the totality of their groceries at a point designed for this effect. This analysis has two main purposes, the first being to provide real life results and an actual useable analysis for Delhaize. This can then be used to improve the Delhaize collect service and build a better customer experience. The second, a personal and intellectual reason, is to understand the functioning of different machine learning tools and be able to apply them to a raw non fictive database. We tackled our problem using an open minded approach. We started by trying to get a feel of the situation by extracting statistics and graphs. Then, when we felt that we had understood the basics, we began brainstorming in order to set up a model at the aggregate level of an INS9 region (Section 4 of the report). This model was essentially about trying to answer the question, “Which of the stores that do not have the collect withdrawal facility should be prioritized for an upgrade?”. We attempted to stay flexible, pursuing the leads that we thought would bring up the most and tried placing ourselves in the shoes of the collect client. At a certain point, this aggregate lead wasn’t yielding enough conclusive results, making us re-evaluate our model from the start. This is when, with the help and discussion with teachers, we moved on to our second and final approach (Section 5). In this approach we went a little deeper by focusing on the individual clients and their purchases. Aside from trying to answer the question stated above, the different steps of the creation of the model also branch out into various useful information. The first model that we made, based on the INS9 region, only provides some very basic depth in terms of usable results. This is still an important part of our report as we managed to make broader conclusions about the data and got the ball rolling. In our second model, however, we managed to get some usable conclusions. Our results show that while the collect clients can not be put into straightforward categories, we can still make reasonably reliable predictions as to whether a client will become a collect client or not, depending on his purchases and other personal features. This in turn can be used for numerous commercial applications (mailing, store prediction, collect site upgrades…)