Denoising Long- and Short-term Interests for Sequential Recommendation

TL;DR

This work tackles noise in sequential recommendation arising from multi-time-scale user interests by introducing LSIDN, which separately denoises long-term and short-term signals. It combines a session-aware long-term encoder (with intra-session Transformer views and inter-session evolution via GRU) and a short-term encoder that uses a homogeneous exchanging augmentation for contrastive learning, followed by an adaptive fusion of $u^L$ and $u^S$ to predict the next item. Key contributions include identifying inter-session behavioral noise, designing session-level long-term denoising, introducing a novel augmentation for short sequences, and demonstrating strong, robust gains on two real-world datasets with extensive ablations. The approach advances robust, multi-scale user modeling in SR and offers practical implications for deploying time-scale-aware recommender systems in noisy real-world data. Overall, LSIDN achieves superior performance and robustness, with open-source code to support reproducibility.

Abstract

User interests can be viewed over different time scales, mainly including stable long-term preferences and changing short-term intentions, and their combination facilitates the comprehensive sequential recommendation. However, existing work that focuses on different time scales of user modeling has ignored the negative effects of different time-scale noise, which hinders capturing actual user interests and cannot be resolved by conventional sequential denoising methods. In this paper, we propose a Long- and Short-term Interest Denoising Network (LSIDN), which employs different encoders and tailored denoising strategies to extract long- and short-term interests, respectively, achieving both comprehensive and robust user modeling. Specifically, we employ a session-level interest extraction and evolution strategy to avoid introducing inter-session behavioral noise into long-term interest modeling; we also adopt contrastive learning equipped with a homogeneous exchanging augmentation to alleviate the impact of unintentional behavioral noise on short-term interest modeling. Results of experiments on two public datasets show that LSIDN consistently outperforms state-of-the-art models and achieves significant robustness.

Figures (6)

• Figure 1: Taking the Taobao dataset as an example, we divide the sequence into sessions according to the time interval $\Delta t$ between two successive items, and set the division threshold $\omega$ = 360min. Note that the diversity of item categories in a session/sequence can be used to reflect the diversity of user behaviors. (a) We randomly selected several sequence instances for qualitative observation. Here, the number on the top half of the circle indicates the item ID, and the number on the bottom half indicates the item category ID. (b) We randomly select two sessions (i.e., $\mathrm{Session}_A$ and $\mathrm{Session}_B$) for each user of the Taobao dataset and then merge them into one (i.e., $\mathrm{Merged\ Session}$). We use two metrics (i.e., entropy and Gini index) zheng2021collaborative to quantitatively evaluate the category diversification of these sessions, where a high entropy or low Gini index implies the great diversity.
• Figure 2: The architecture of our proposed model LSIDN.
• Figure 3: Illustration of our proposed augmentation method.
• Figure 4: Performance comparison of long- and short-term interest models.
• Figure 5: Robustness analysis of different methods. The bar represents the evaluation metrics, and the line represents the percentage of performance decrease, where a smaller drop rate indicates better noise resistance.