Beyond Uniform Modeling: A Dual-View MLP Framework for Trend and Diversity Disentanglement in Multibehavior Recommendation

Abstract

In real-world recommendation systems, user interactions contain behavioral information. Multibehavior sequential recommendation focuses on predicting the next item a user will interact with under a specific behavior. Incorporating behavioral information introduces diverse interaction patterns, which we treat as auxiliary signals to supplement trend interest modeling (e.g., primary interests, low-frequency interests, and noise). However, existing multibehavior recommendation methods perform unified modeling without distinguishing the interest categories reflected by different data. Capturing both interest trends and dispersed diversity inherently presents a conflicting objective. Hence, we propose a novel approach that separately models these aspects before integrating them. To realize this vision, we introduce Disentangled Interest Perspective for Multibehavior Sequential Recommendation (DIPRec), which separately captures interest tendencies and diversity before performing a unified integration. The framework begins with a mask generation module that partitions the data, employing multilayer perceptron (MLP)-based modules—specifically, the heterogeneous interest trend module (HITM) and the auxiliary behavior intent perception module (ABIP) to extract user interest tendencies from heterogeneous and auxiliary behavioral sequences. Concurrently, it leverages MLP and max-pooling layers to extract supplementary diversity signals from the separated data. Finally, the model integrates these two aspects to generate recommendations. To assess the effectiveness of our proposed DIPRec model, we conducted thorough experiments on five public datasets. The results consistently demonstrate that DIPRec surpasses current state-of-the-art approaches across a variety of standard evaluation metrics.