Disentangle interest trend and diversity for sequential recommendation

Based on historical behaviors, sequential recommendation endeavors to predict what a user prefers next. The recent efforts are mainly devoted to modeling the user's interests evolution process or mining multi-interests for recommendation. However, it is largely overlooked that the interest trend (i.e., the evolution of the main interest) and the interest diversity (i.e., the scattered potential interests) could complement each other for better performance. Specifically, the interest trend reveals the user's basic interest and its evolution, which is satisfied by similarity recommendations. Nevertheless, interest diversity covers the various interests caused by some external environmental influence, e.g., fashion trends and advertisements, exploring users' potential interests or interest diversity will facilitate the model for diversity and serendipity recommendation. In a way, these two factors have conflicting aims, we argue that they should be disentangled in modeling first and recombined when making personalized recommendation. To instantiate this idea, we propose a simple yet effective model, dubbed TEDDY (disentangles interest trend and diversity), which disentangles and then jointly models the aforementioned two factors under a unified framework. Particularly, in TEDDY, an adaptive masking mechanism is first introduced to split the user's historical items into two parts revealing her major interest trend and scattered interest diversity respectively for interests disentanglement. Then, a temporal convolutional network (TCN) is utilized to capture the evolution process of the user's major interest trend. For the scattered interest diversity modeling, we further choose to apply Multilayer Perceptron (MLP) layers with max-pooling mechanism to extract the significant or dominant preference signals. The predicted scores generated by these two modules are aggregated together to integrate both interest trend and interest diversity for the final recommendation. Extensive experiments over four public datasets demonstrate the superiority of our proposed TEDDY against a series of SOTA alternatives on the benchmark metrics.