Recommendation Systems with Causal Inference-based Regularization

Nakul Ravindra Padalkar, Victor De Lima

Abstract: Description Recommender systems have become integral to modern society, facilitating personalized experiences across various platforms by allowing the offering of tailored content, products, and services. Current approaches for building such systems employ neural networks that overfit the training data and fail to generalize when tested on previously unseen data. Furthermore, recommender systems are fundamentally causal inference problems since they seek to forecast conditional relationships that describe a user's response to specific advertisements or products. In this paper, we comprehensively examine the matrix factorization model for recommendation systems using the MovieLens 25M dataset. We demonstrate the limitations of standard collaborative filtering and motivate the need for more suitable regularization. This research lays the foundation for future work incorporating causal inference into recommendation systems, which has garnered increasing support in recent literature for its potential to improve model performance.