Linear Item-Item Model with Neural Knowledge for Session-based Recommendation

TL;DR

The paper tackles the gap between neural sequence models and linear co-occurrence models in session-based recommendation by proposing LINK, a hybrid linear framework composed of LIS and NIT. LIS performs self-distillation to refine item-item similarity and propagate intra- to inter-session signals, while NIT distills neural knowledge into a linear transition matrix derived from an off-the-shelf neural model. The two components are fused into a closed-form solution $B^{LINK}$, enabling fast inference for new sessions via a simple matrix multiplication, and the framework is model-agnostic with respect to the neural teacher. Empirical results across six real-world datasets show LINK achieving up to 14.78% gains in Recall@20 and 11.04% in MRR@20 compared to strong linear baselines, while offering up to 813x lower inference FLOPs than neural models, indicating a practical and scalable path for high-performance, efficient SBR.

Abstract

Session-based recommendation (SBR) aims to predict users' subsequent actions by modeling short-term interactions within sessions. Existing neural models primarily focus on capturing complex dependencies for sequential item transitions. As an alternative solution, linear item-item models mainly identify strong co-occurrence patterns across items and support faster inference speed. Although each paradigm has been actively studied in SBR, their fundamental differences in capturing item relationships and how to bridge these distinct modeling paradigms effectively remain unexplored. In this paper, we propose a novel SBR model, namely Linear Item-Item model with Neural Knowledge (LINK), which integrates both types of knowledge into a unified linear framework. Specifically, we design two specialized components of LINK: (i) Linear knowledge-enhanced Item-item Similarity model (LIS), which refines the item similarity correlation via self-distillation, and (ii) Neural knowledge-enhanced Item-item Transition model (NIT), which seamlessly incorporates complicated neural knowledge distilled from the off-the-shelf neural model. Extensive experiments demonstrate that LINK outperforms state-of-the-art linear SBR models across six real-world datasets, achieving improvements of up to 14.78% and 11.04% in Recall@20 and MRR@20 while showing up to 813x fewer inference FLOPs. Our code is available at https://github.com/jin530/LINK.

Figures (4)

• Figure 1: Comparison of how linear and neural models learn item relationships. The analysis is performed on consecutive item pairs (previous item $\rightarrow$ next item) from validation sessions in the Diginetica dataset. The x- and y-axis indicate the popularity of the previous and next items, respectively. The color represents the model's predicted correlation score. For simplicity, we linearly scale the predicted score into $[0, 1]$.
• Figure 2: Overall training framework of LINK, consisting of Linear Knowledge-enhanced Item-Item Similarity Model (LIS) and Neural Knowledge-enhanced Item-Item Transition Model (NIT).
• Figure 3: Performance comparison over the Recall@20 and the log-scaled number of floating point operations (FLOPs) for inference on Diginetica and Yoochoose datasets. The numbers in parentheses indicate the relative FLOPs compared to LINK.
• Figure 4: Performance comparison of representative neural and linear SBR models on Head and Tail target items. Head comprises target items within the top 20% of popular items, while Tail consists of the bottom 80% of less popular items.