Abstract
Session-based recommendation (SBR) has drawn increasingly research attention in recent years, due to its great practical value by only exploiting the limited user behavior history in the current session. Existing methods typically learn the session embedding at the item level, namely, aggregating the embeddings of items with or without the attention weights assigned to items. However, they ignore the fact that a user's intent on adopting an item is driven by certain factors of the item (e.g., the leading actors of a movie). In other words, they have not explored finer-granularity interests of users at the factor level to generate the session embedding, leading to sub-optimal performance. To address the problem, we propose a novel method called Disentangled Graph Neural Network (Disen-GNN) to capture the session purpose with the consideration of factor-level attention on each item. Specifically, we first employ the disentangled learning technique to cast item embeddings into the embedding of multiple factors, and then use the gated graph neural network (GGNN) to learn the embedding factor-wisely based on the item adjacent similarity matrix computed for each factor. Moreover, the distance correlation is adopted to enhance the independence between each pair of factors. After representing each item with independent factors, an attention mechanism is designed to learn user intent to different factors of each item in the session. The session embedding is then generated by aggregating the item embeddings with attention weights of each item's factors. To this end, our model takes user intents at the factor level into account to infer the user purpose in a session. Extensive experiments on three benchmark datasets demonstrate the superiority of our method over existing methods.
Introduction
Session-based recommendation (SBR) represents a significant challenge in modern e-commerce and content platforms. Unlike traditional recommendation systems that rely on long-term user profiles, SBR must infer user intent from a brief, anonymous sequence of interactions within a single session. This "cold start" scenario requires models to extract maximum information from limited data.
A critical limitation of current SBR models is their treatment of items as holistic, indivisible entities. In reality, a user's decision to click on an item is often motivated by specific latent factors—such as the brand of a shirt, the genre of a movie, or the artist of a song. By aggregating item embeddings without considering these fine-grained attributes, existing methods fail to capture the nuanced factors driving user behavior, resulting in sub-optimal recommendation accuracy.
Methodology
To overcome these limitations, we introduce Disen-GNN (Disentangled Graph Neural Network). The core philosophy of Disen-GNN is to decompose item representations into multiple independent factors, allowing the model to attend to specific user interests at a finer granularity.
The architecture consists of several key stages:
- Disentangled Item Embedding: Each item embedding is partitioned into $K$ independent components, each representing a distinct latent factor.
- Factor-wise Gated Graph Neural Networks (GGNN): Information is propagated across the session graph independently for each factor. By computing a separate similarity matrix for each factor, the model learns how items relate based on specific attributes rather than general similarity.
- Distance Correlation Regularization: To ensure that the $K$ factors remain distinct and non-redundant, a distance correlation loss is applied, enforcing statistical independence between factor pairs.
- Residual Attention Mechanism (RAM): To mitigate the over-smoothing problem inherent in deep GNNs—where node features become too similar after multiple layers—the RAM module preserves the unique identity of the original item embeddings.
- Factor-level Attention: The final session embedding is constructed using an attention mechanism that weights the importance of different factors across all items in the session, using the last interacted item to anchor the user's current intent.
Experimental Evaluation
Disen-GNN was evaluated against state-of-the-art baselines (e.g., SR-GNN, TAGNN) on three benchmark datasets: Diginetica (e-commerce), Yoochoose 1/64 (clickstreams), and Nowplaying (music).
Key Results
The proposed model consistently outperformed all baselines across all metrics. On the Diginetica dataset, Disen-GNN achieved a Precision@20 of 53.79% and an MRR@20 of 18.99%, representing a 4-5% improvement over the previous best-performing model, TAGNN. The performance gains were even more pronounced on the Nowplaying dataset, where Disen-GNN showed a ~16% relative improvement in Precision@20. These results indicate that Disen-GNN not only identifies relevant items more accurately but also ranks them higher in the recommendation list.
Insights and Ablation Studies
- Granularity Matters: Experiments varying the number of factors ($K$) showed that $K=1$ (holistic items) consistently performed worst. Optimal $K$ values varied by domain: $K=5$ for general e-commerce (Diginetica) and $K=12$ for music (Nowplaying), suggesting that music preferences are driven by a more diverse set of latent factors.
- Independence is Crucial: Removing the distance correlation regularization ($λ=0$) led to a significant drop in performance, confirming that the independence of factors is vital for robust feature learning.
- Depth and Over-smoothing: The Residual Attention Mechanism proved essential for maintaining performance as the number of GNN layers increased, allowing the model to capture higher-order relationships without losing item-specific signals.
Conclusion
Disen-GNN demonstrates that disentangling item embeddings into latent factors significantly enhances session-based recommendation. By modeling factor-level user intent and ensuring factor independence, the framework provides a more nuanced understanding of anonymous user behavior. Future work will focus on reducing the computational overhead of distance correlation and incorporating global session patterns to further refine recommendation quality.
Summary updated