Prediction of Online Upselling

Challenge

Analyzing customer data on websites and portals is a typical application for data science, statistics, and machine learning. Many online user actions leave digital traces, database entries, and connections between entities. Our client, an international online company, offers a paid membership with various benefits alongside a free plan. The project's challenge was to use the many different onsite data to develop a model for targeted identification of purchase-ready customers.

Approach

A deep learning model was used to model upselling propensity. Deep learning is particularly well-suited for classification problems when there is little or no prior information about the causal relationships between potential influencing factors and the outcome (in this case: upselling). The multi-layer architecture of the neural network allows for the creation of features that consist of various interactions of influencing factors. These features are then used to generate an optimal forecast. As part of the project, we trained an appropriate model.

Result

The deep learning model performed significantly better than the existing upselling model. Additionally, an infrastructure was provided via the H2O platform, on which the model is based, allowing the finalized model to evaluate any user for upselling potential at the push of a button.