Customer return
predictions
Given the order of a customer in an online shop, can we predict which items will be returned later?
Motivation
Return rates in E-commerce are significantly higher than in retail. Especially when buying clothes online, customers expect free return shipping. This comes with immense economic as well as ecological costs.
Therefore, online retailers are interested in predicting customer returns at the time of their order. This enables them to better plan ahead. It also creates the possibility to subtly nudge users to not order a specific item they will return with a high probability.
Data
The data originates from real orders in an online fashion store. It was provided by Humboldt University in the context of this course project. It contains information about the ordered items and their binary labels describe if the customer returned a specific item or not. It also contains user IDs that were generated for this dataset and that are not traceable to the actual customers. The dataset is balanced such that about half of the listed items were returned.
Approach
Most of the work in this project encompasses data exploration, data cleaning and feature engineering. Afterwards, the data will be classified by experimenting with Logistic Regression and a Random Forest Classifier, and then using the model that after parameter tuning minimizes our loss function on test data.
This problem has an asymmetric loss function that is based on the following idea. If we incorrectly predict that a customer will return an item, we nudge them to not buy it, resulting in a possible loss of profit. If we fail to predict that a customer will return an item, the store has to pay shipping costs as well as a fee for reconditioning the item.
To conveniently create features I wrote a custom scikit-learn transformer. This makes the main part of the code in the notebook below easier to read. It also enables us to effectively use it in pipelines.
Learnings
This project is a good example on how important thorough data exploration and domain knowledge are in Data Science. It also showcases how categories like color and size, whose naming and interpretation can be subjective, are difficult to process in a model. Summing it up in one sentence: Understand your data and what it presents, before you even think about applying a model to it.