Expressive user embedding from churn and recommendation multi-task learning

In this paper, we present a Multi-Task model for Recommendation and Churn prediction (MT) in the retail banking industry. The model leverages a hard parameter-sharing framework and consists of a shared multi-stack encoder with multi-head self-attention and two fully connected task heads. It is trained to achieve two multi-class classification tasks: predicting product churn and identifying the next-best products (NBP) for users, individually. Our experiments demonstrate the superiority of the multi-task model compared to its single-task versions, reaching top-1 precision at 78.1% and 77.6%, for churn and NBP prediction respectively. Moreover, we find that the model learns a coherent and expressive high-level representation reflecting user intentions related to both tasks. There is a clear separation between users with acquisitions and users with churn. In addition, acquirers are more tightly clustered compared to the churners. The gradual separability of churning and acquiring users, who diverge in intent, is a desirable property. It provides a basis for model explainability, critical to industry adoption, and also enables other downstream applications. These potential additional benefits, beyond reducing customer attrition and increasing product use-two primary concerns of businesses, make such a model even more valuable.