SPGL: Enhancing Session-based Recommendation with Single Positive Graph Learning

TL;DR

This work tackles data sparsity in session-based recommendations by introducing SPGL, which combines graph convolutional networks with a single positive optimization loss to learn robust global-item and session representations. SPGL builds a directed 3-hop global item graph and employs an intent extractor that incorporates reverse positional information, enabling effective session representation learning. The approach achieves strong results on Tmall, RetailRocket, and Diginetica, with ablations showing the key impact of the single positive loss and temporal information, while preserving simplicity compared to multi-view contrastive setups. Overall, SPGL provides a lightweight yet powerful alternative to contrastive sampling, delivering practical improvements for real-world session-based recommendation tasks, and the code is publicly available.

Abstract

Session-based recommendation seeks to forecast the next item a user will be interested in, based on their interaction sequences. Due to limited interaction data, session-based recommendation faces the challenge of limited data availability. Traditional methods enhance feature learning by constructing complex models to generate positive and negative samples. This paper proposes a session-based recommendation model using Single Positive optimization loss and Graph Learning (SPGL) to deal with the problem of data sparsity, high model complexity and weak transferability. SPGL utilizes graph convolutional networks to generate global item representations and batch session representations, effectively capturing intrinsic relationships between items. The use of single positive optimization loss improves uniformity of item representations, thereby enhancing recommendation accuracy. In the intent extractor, SPGL considers the hop count of the adjacency matrix when constructing the directed global graph to fully integrate spatial information. It also takes into account the reverse positional information of items when constructing session representations to incorporate temporal information. Comparative experiments across three benchmark datasets, Tmall, RetailRocket and Diginetica, demonstrate the model's effectiveness. The source code can be accessed on https://github.com/liang-tian-tian/SPGL .

Figures (7)

• Figure 1: The design of the SPGL framework. (A) A global directed graph with 3-hop neighbors is formulated based on user interactions. (B) Session representation learning and item embeddings are derived using graph convolutional networks and attention mechanisms. Single positive optimization enhances item representation, and the prediction layer computes relevance scores for candidate items.
• Figure 2: Construction of the 3-hop directed global item graph. We first construct the directed global graph across all sessions. Subsequently, we add $\epsilon$-neighbors weighted edges to more precisely capture information. The example illustrates the adjacency and weight calculations for a 3-hop directed subgraph centered at $v_{2}$.
• Figure 3: An illustration of the principle of single positive optimization. (a) Without single positive optimization, similar items are closely aligned in the feature space. (b) With single positive optimization, the item representation space achieves better uniformity, providing a valuable complement to cross-entropy loss by enhancing the dispersion of item features.
• Figure 4: Impact of the reverse positional information
• Figure 5: Impact of the hop range