Abstract
Traditional robust recommendation methods view atypical user-item interactions as noise and aim to reduce their impact with some kind of noise filtering technique, which often suffers from two challenges. First, in real world, atypical interactions may signal users' temporary interest different from their general preference. Therefore, simply filtering out the atypical interactions as noise may be inappropriate and degrade the personalization of recommendations. Second, it is hard to acquire the temporary interest since there are no explicit supervision signals to indicate whether an interaction is atypical or not. To address this challenges, we propose a novel model called Temporary Interest Aware Recommendation (TIARec), which can distinguish atypical interactions from normal ones without supervision and capture the temporary interest as well as the general preference of users. Particularly, we propose a reinforcement learning framework containing a recommender agent and an auxiliary classifier agent, which are jointly trained with the objective of maximizing the cumulative return of the recommendations made by the recommender agent. During the joint training process, the classifier agent can judge whether the interaction with an item recommended by the recommender agent is atypical, and the knowledge about learning temporary interest from atypical interactions can be transferred to the recommender agent, which makes the recommender agent able to alone make recommendations that balance the general preference and temporary interest of users. At last, the experiments conducted on real world datasets verify the effectiveness of TIARec.
Introduction
Recommender systems often struggle when users exhibit behavior that deviates from their historical patterns. For instance, a user who typically watches romantic comedies might suddenly engage with action movies during a holiday weekend. Traditional "robust" recommendation models treat such atypical interactions as noise, employing filtering techniques to minimize their influence on the user profile.
However, this approach has significant drawbacks. First, it ignores the fact that atypical interactions often represent genuine, albeit temporary, interests. Filtering them out can lead to a loss of personalization and a failure to meet the user's immediate needs. Second, identifying which interactions are "noise" and which are "temporary interests" is difficult due to the lack of explicit supervision signals.
The TIARec Framework
To overcome these limitations, we introduce Temporary Interest Aware Recommendation (TIARec), a multi-agent reinforcement learning (RL) framework designed to distinguish between general preferences and temporary interests without explicit labels.
As illustrated in the architecture below, TIARec employs a dual-agent system:
- Recommender Agent ($\Pi_r$): Generates item suggestions by weighing historical interactions using an attention mechanism.
- Auxiliary Classifier Agent ($\Pi_c$): Acts as a supervisor that categorizes interactions into normal (general preference) or atypical (temporary interest) sets.
- Critic ($Q$): Evaluates the joint actions of both agents using the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm to maximize long-term cumulative returns.
The core innovation lies in the joint training process. The classifier agent "teaches" the recommender agent by identifying atypical interests that lead to high rewards. This knowledge is transferred to the recommender agent, enabling it to balance long-term preferences and short-term interests independently during inference.
Experimental Evaluation
We evaluated TIARec on three diverse real-world datasets: Movielens-1M (movie ratings), Beauty (Amazon product reviews), and Tianchi (TMALL e-commerce data). These datasets provide a broad range of user behaviors and interaction densities to test the model's robustness and adaptability.
Key Results and Discussion
TIARec consistently outperformed state-of-the-art robust models, including CDAE and APR, across multiple metrics such as Hit Rate (HR) and Normalized Discounted Cumulative Gain (NDCG).
Key findings include:
- Superior Performance: On the Movielens-1M dataset, TIARec achieved a 22.65% improvement in HR@5 and a 39.80% improvement in NDCG@10 compared to the strongest baseline.
- Effective Interest Balancing: The model successfully captures "pivots" in user behavior. While baselines often fail to adapt to sudden shifts in interest, TIARec identifies these temporary trends and incorporates them into its recommendations without discarding the user's long-term profile.
- Robustness to Noise: TIARec maintains high performance even when simulated atypical behavior is increased, demonstrating its ability to distinguish meaningful temporary interests from truly random noise.
Conclusion
TIARec represents a paradigm shift in robust recommendation by treating atypical behavior as a valuable signal rather than mere noise. By leveraging a multi-agent reinforcement learning framework, the model learns to navigate the ambiguity of user interests without manual labeling. Future work will explore incorporating additional contextual signals, such as temporal and spatial data, to further refine the classification of temporary interests.
Summary updated