TL;DR
Disen-GNN is a novel recommendation framework that breaks down items into independent "factors" (like style, brand, or actor) to better understand anonymous user intent in a single session. By using factor-wise Graph Neural Networks and a specialized attention mechanism, it achieves state-of-the-art performance on major recommendation benchmarks.
What Problem Does This Paper Address?
Session-based recommendation (SBR) is the "cold start" challenge of the e-commerce world. You have an anonymous user and only a handful of clicks in their current session to figure out what they want next.
The authors identify a major gap in existing SBR models: they treat items as holistic entities. In reality, a user's interest is often driven by specific factors of an item. For example, a user might click on a movie because of a specific actor, or a shirt because of its color. Existing methods aggregate item embeddings into a "session embedding" without considering these fine-grained, factor-level interests, leading to sub-optimal recommendations.
Methods
The authors propose Disen-GNN (Disentangled Graph Neural Network), which introduces several novel components:
- Disentangled Item Embedding: Instead of one large vector, an item's embedding is split into $K$ independent chunks, each representing a latent factor.
- Factor-wise GGNN: The model uses Gated Graph Neural Networks (GGNN) to propagate information across the session graph. Crucially, it computes a separate similarity matrix for each factor, allowing the model to learn how items relate to one another based on specific attributes.
- Distance Correlation Regularization: To ensure the $K$ factors are truly independent and not redundant, the authors apply a distance correlation loss function during training.
- Residual Attention Mechanism (RAM): To combat the "over-smoothing" problem common in deep GNNs (where node representations become indistinguishable), a RAM module is used to preserve the unique features of the original item embeddings.
- Factor-level Attention: A final attention layer determines which factors of which items in the session are most important for predicting the next click, using the last clicked item as a proxy for "local" intent.
Datasets
The model was evaluated on three widely used real-world benchmarks:
- Diginetica: Transaction data from the CIKM Cup 2016.
- Yoochoose 1/64: A subset of user click behaviors from an e-commerce website.
- Nowplaying: Music listening behaviors extracted from Twitter.
Key Insights and Ideas
- The "Aha" of Disentanglement: The experiments show that $K=1$ (treating items holistically) consistently yields the worst results. This confirms that breaking items into factors is essential for capturing complex user intent.
- Independence Matters: Setting the regularization coefficient $\lambda$ to zero (removing the independence constraint) significantly hurts performance. This proves that forcing factors to be statistically independent helps the model learn more robust features.
- Solving Over-smoothing: The authors found that without their Residual Attention Mechanism, performance dropped sharply as they added more GNN layers. RAM allows the model to "go deeper" and exploit higher-order neighbor information without losing the signal of the individual items.
- Dataset Specificity: Different datasets require different levels of granularity. Diginetica performed best with $K=5$ factors, while the music-based Nowplaying dataset peaked at $K=12$, suggesting music preferences are driven by a wider variety of latent factors than general e-commerce clicks.
Key Results
Disen-GNN outperformed all baselines, including popular models like SR-GNN and TAGNN:
- Diginetica: Achieved a P@20 of 53.79% and MRR@20 of 18.99%, roughly a 4-5% improvement over the previous best (TAGNN).
- Nowplaying: Showed a massive jump, with a P@20 of 22.22% compared to TAGNN's 19.02% (a ~16% relative improvement).
- Ranking Quality: The consistently higher MRR (Mean Reciprocal Rank) scores indicate that Disen-GNN is not just getting the right items in the top 20, but is placing the actual next item higher up in the list.
Limitations and Future Work
- Computational Complexity: The distance correlation used for disentanglement is computationally expensive. The authors could only afford to apply it to the first layer of the network.
- Memory Constraints: On the large Yoochoose dataset, hardware limitations prevented the authors from testing more than 5 factors, potentially leaving some performance on the table.
- Future Directions: The authors suggest exploring "global" information (patterns across all sessions in a dataset) and external context to further refine the factor-level interest modeling.
Summary updated