Bid Farewell to Seesaw: Towards Accurate Long-tail Session-based Recommendation via Dual Constraints of Hybrid Intents

TL;DR

This work tackles the long-tail bias in session-based recommendation by identifying and suppressing session-irrelevant tail noise. It introduces HID, a plug-and-play framework that combines Hybrid Intent Learning (via attribute-aware spectral clustering to form hybrid intents) with an Intent Constraint Loss that jointly optimizes for long-tail diversity and predictive accuracy. The framework defines target and noise intents to supervise embedding learning and proves two theoretical results that justify the efficiency and effectiveness of the constraints. Extensive experiments across multiple SBR models and datasets show HID achieves state-of-the-art long-tail performance without sacrificing accuracy, establishing a practical, generalizable approach for accurate long-tail SBR.

Abstract

Session-based recommendation (SBR) aims to predict anonymous users' next interaction based on their interaction sessions. In the practical recommendation scenario, low-exposure items constitute the majority of interactions, creating a long-tail distribution that severely compromises recommendation diversity. Existing approaches attempt to address this issue by promoting tail items but incur accuracy degradation, exhibiting a "see-saw" effect between long-tail and accuracy performance. We attribute such conflict to session-irrelevant noise within the tail items, which existing long-tail approaches fail to identify and constrain effectively. To resolve this fundamental conflict, we propose \textbf{HID} (\textbf{H}ybrid \textbf{I}ntent-based \textbf{D}ual Constraint Framework), a plug-and-play framework that transforms the conventional "see-saw" into "win-win" through introducing the hybrid intent-based dual constraints for both long-tail and accuracy. Two key innovations are incorporated in this framework: (i) \textit{Hybrid Intent Learning}, where we reformulate the intent extraction strategies by employing attribute-aware spectral clustering to reconstruct the item-to-intent mapping. Furthermore, discrimination of session-irrelevant noise is achieved through the assignment of the target and noise intents to each session. (ii) \textit{Intent Constraint Loss}, which incorporates two novel constraint paradigms regarding the \textit{diversity} and \textit{accuracy} to regulate the representation learning process of both items and sessions. These two objectives are unified into a single training loss through rigorous theoretical derivation. Extensive experiments across multiple SBR models and datasets demonstrate that HID can enhance both long-tail performance and recommendation accuracy, establishing new state-of-the-art performance in long-tail recommender systems.

Figures (5)

• Figure 1: Comparison between the our proposed HID and previous work. (a) illustrates the design of HID, where t. and n. denotes target and noise items for session $S^u$, respectively; (c) and (d) demonstrate the frameworks of previous long-tail approaches. (b) evaluates the accuracy (i.e., HR@20) and long-tail performance (i.e., tCov@20) of the base SBR model GRU4Rec Hidasi16gru4rec and GRU4Rec + long-tail approaches on Tmall dataset.
• Figure 2: The demonstration of: (a) Hybrid Intent: Step 1 groups items by shared attributes (e.g., food) as the preliminary intents; Step 2 combines attributes with high co-occurrence (e.g., food + pot) to form the hybrid intents (e.g., cooking). (b) Intent Assignment: Assigns target (relevant) and noise (irrelevant) hybrid intents to anonymous sessions.
• Figure 3: The overall architecture of $SBR~model$ (left) + $HID$ (right). The Hybrid Intent Learning module first assigns items to $k$ preliminary intents, and then further divides them into $n$ hybrid intents $C$ based on the topological relationships in the preliminary intent graph. After refining the hybrid intents, the intent constraint loss is introduced to regulate the learning process of session embedding $\textbf{S}^u$.
• Figure 4: The changes in accuracy (HR@20) and long-tail (tCov@20) metrics with the increase of scale $\epsilon$. The model is SRGNN+HID.
• Figure 5: The changes in accuracy (HR@20) and long-tail (tCov@20) metrics with the increase of clusters $n$. The model is SRGNN+HID.