Abstract
Sequential Recommender Systems (SRS) typically treat all historical user interactions as equally significant. However, real-world data is often noisy, containing "mis-clicks" or interactions driven by external trends rather than intrinsic user interest. This paper introduces AutoSAM, an automatic sampling framework that utilizes reinforcement learning to identify and retain only the most informative user behaviors. By filtering out noise, AutoSAM improves both the accuracy and computational efficiency of recommendations. Experimental results across multiple datasets demonstrate that AutoSAM consistently outperforms state-of-the-art baselines and is compatible with various recommender architectures.
Introduction
The primary goal of Sequential Recommender Systems (SRS) is to predict a user's future actions based on their historical behavior. A significant limitation of standard SRS models is the assumption that all historical interactions are "uniformly positive." In practice, user behavior is often stochastic and noisy; interactions may stem from accidental clicks, idle curiosity, or "herd mentality" rather than genuine intent.
Treating these noisy interactions with the same weight as high-intent actions leads to skewed user representations. While some existing noise-filtering methods exist, they are often "target-dependent," requiring a known candidate item to filter history during the ranking stage. This makes them incompatible with the efficient generative training losses used in modern recommenders. AutoSAM addresses this by providing a target-independent, intelligent sampling layer.
Methodology
AutoSAM introduces a lightweight "sampler" layer positioned before the core recommendation model. This sampler, composed of a Transformer block and an MLP, evaluates each item in a user's history to decide whether it should be kept or discarded.
Reinforcement Learning and Multi-Objective Rewards
Since the sampling decision is a discrete operation, the framework is trained using policy gradient methods. The sampler is optimized based on two primary reward signals:
- Future Prediction: Measures how much a retained item contributes to predicting the user's subsequent actions.
- Sequence Perplexity: Inspired by NLP, this reward encourages the selection of items that fit the sequence's context while penalizing outliers (high perplexity).
The sampler and the backbone recommendation model are trained end-to-end, allowing the sampler to adapt as the recommender's predictive capabilities improve.
Experimental Evaluation
The authors evaluated AutoSAM using four real-world datasets: Tmall, Alipay, Yelp, and Amazon Book. The framework was tested as a "plug-in" across different backbone architectures, including GRU4Rec (RNN), NextitNet (CNN), and SASRec (Transformer).
Key Results
AutoSAM consistently achieved superior performance compared to state-of-the-art baselines. Notably, on the Amazon Book dataset, it improved Recall@10 and NDCG@10 by over 14%. Qualitative analysis confirmed that AutoSAM effectively prioritizes high-intent actions, such as purchases, over lower-intent clicks.
Beyond accuracy, AutoSAM enhances efficiency. By reducing the length of input sequences through intelligent sampling, it significantly lowers the computational cost (MFLOPs) during inference without sacrificing—and often improving—recommendation quality.
Conclusion
AutoSAM shifts the focus from model architecture to data quality in Sequential Recommender Systems. By automatically filtering noisy user interactions through a reinforcement learning-based sampler, the framework achieves significant gains in both accuracy and efficiency. Future work will explore handling cold-start scenarios for new users and automating hyperparameter tuning for reward sensitivity.
Summary updated