[4D1-GS-2] Machine learning: Uncertainty / Information visualization

[4D1-GS-2-02] Twin-G NeuMF: An Enhanced Neural Matrix Factorization Model Robust to Noise

〇Shogo Chomei¹, Ryuta Matsuoka¹, Naru Shimizu¹, You Tensyou², Masayuki Goto¹ (1. Waseda University, 2. Keio University)

Keywords:Recommendation Model, Collaborative Filtering, Implicit Feedback, Noise Data, Negative Sampling

In recent years, there has been a growing demand for recommendation models that provide suggestions tailored to preferences of user from a vast array of items on e-commerce websites and other platforms. In particular, collaborative filtering which utilizes implicit feedback derived from user interactions (such as purchases and browsing) plays a crucial role in supporting current recommendation systems, and one of these techniques is Neural Matrix Factorization (NeuMF). NeuMF is a model that integrates linear and non-linear models, enabling the representation of complex structures. However, implicit feedback is susceptible to noise due to the ease of interaction and the fact that users do not directly provide ratings. Since NeuMF is a non-linear model, it is prone to overfitting to noise, such as accidental click interactions. Therefore, in this study, we propose Twin-G NeuMF, which adds an additional linear mechanism to the conventional NeuMF. From experiments, we demonstrate the effectiveness of the proposed method, particularly in scenarios with high levels of noise. Additionally, the proposed method demonstrates that it can mitigate the impact of the ease with which negative sampling overlaps, due to a large proportion of the data having observed interactions.