From collaborative filtering to deep learning: Advancing recommender systems with longitudinal data in the financial services industry

Recommender systems (RS) are highly relevant for multiple domains, allowing to construct personalized suggestions for consumers. Previous studies have strongly focused on collaborative filtering approaches, but the inclusion of longitudinal data (LD) has received limited attention. To address this gap, we investigate the impact of incorporating LD for recommendations, comparing traditional collaborative filtering approaches, multi-label classifier (MLC) algorithms, and a deep learning model (DL) in the form of gated recurrent units (GRU). Additional analysis for the best performing model is provided through SHapley Additive exPlanations (SHAP), to uncover relations between the different recommended products and features. Thus, this article contributes to operational research literature by (1) comparing several MLC techniques and RS, including state-of-the-art DL models in a real-life scenario, (2) the comparison of various featurization techniques to assess the impact of incorporating LD on MLC performance, (3) the evaluation of LD as sequential input through the use of DL models, (4) offering interpretable model insights to improve the understanding of RS with LD. The results uncover that DL models are capable of extracting information from longitudinal features for overall higher and statistically significant performance. Further, SHAP values reveal that LD has the higher impact on model output and managerial relevant temporal patterns emerge across product categories.