Abstract
Session-based recommendation (SBR) is widely used in e-commerce and streaming services, with the task of performing real-time recommendations based on short-term anonymous user history data. Most existing SBR frameworks follow the pattern of learning a single representation for a specific session, which makes it difficult to capture potential multiple interests, thus preventing discriminative recommendations. Multi-Interest learning has emerged as an effective approach for addressing this issue on sequential data in recent years. However, the current Multi-Interest frameworks perform poorly on session data because they may generate excessive interests. To address these issues, we proposed a model named Dynamic Multi-Interest Graph Neural Network (DMI-GNN), which introduces the Multi-Interest learning framework into SBR and refines it by proposing a multiple positional patterns (MPP) learning method and a Dynamic Multi-Interest (DMI) regularization. Specifically, the MPP learning layer ensures the model obtains representations with different positional information for sessions. The DMI regularization, on the other hand, mitigates the influence of excessive interests. Experiments on three benchmark datasets demonstrate that our methods achieve better performance on different metrics.
Introduction
Session-Based Recommendation (SBR) aims to predict a user's next action based on a brief, anonymous sequence of interactions, such as a single shopping trip. A major limitation of existing models is the "representation bottleneck": they typically compress an entire session into a single vector. However, a user's session often reflects multiple distinct interests. For instance, a user might browse for a smartphone, then a protective case, and finally headphones within the same session. Forcing these diverse interests into a single embedding leads to "blurry" recommendations that fail to capture the user's specific intent.
While multi-interest models have been successful for long-term user histories, they struggle with SBR. Because sessions are often very short (averaging 5–10 items), standard multi-interest models tend to generate redundant or low-quality interests, which can actually degrade performance. The core challenge lies in capturing multiple interests when they exist while remaining focused when the session is too short to support them.
Methodology: The DMI-GNN Framework
The proposed Dynamic Multi-Interest Graph Neural Network (DMI-GNN) addresses these challenges through a specialized architecture designed to handle the unique constraints of session data.
1. Graph Context Learning Layer
The model first employs a Graph Neural Network (GNN) to extract relationships between items. This layer captures both global transitions across all sessions and local transitions within the current session, providing a rich contextual embedding for each item.
2. Multiple Positional Patterns (MPP)
Unlike standard models that use a single set of positional encodings, DMI-GNN learns multiple positional patterns. This allows the model to recognize that the importance of an item's position might vary depending on the specific interest being extracted—for example, distinguishing between a targeted search and casual browsing.
3. Multi-Interest Extractor
A multi-head self-attention mechanism is used to generate a fixed number ($U$) of interest vectors. This component allows the model to attend to different subsets of items in the session simultaneously.
4. Dynamic Multi-Interest (DMI) Regularization
The core innovation is the DMI regularization term, which uses session length as a prior to control model complexity. Instead of a fixed number of interests, the model dynamically adjusts the penalty for similarity between interest vectors:
- Short Sessions: The penalty is low, allowing interest vectors to collapse into a similar state, effectively acting as a single interest.
- Long Sessions: The penalty increases, forcing the model to differentiate between interests to capture the user's evolving needs.
Experiments and Results
DMI-GNN was evaluated on three standard real-world benchmarks: Tmall (shopping logs), RetailRocket (browsing data), and LastFM (music listening).
Performance Gains
DMI-GNN outperformed nine state-of-the-art baselines across all metrics. Notable results include:
- Accuracy: On the Tmall dataset, DMI-GNN achieved an HR@20 of 40.26, significantly surpassing previous bests.
- Diversity: The model showed a massive improvement in coverage. On Tmall, coverage jumped to 94.00%, indicating a much better ability to recommend a wide variety of items from the catalog.
- Ablation Studies: Removing either the DMI regularization or the MPP layer led to consistent performance drops, confirming that the dynamic nature of the model is critical to its success.
Conclusion and Future Work
DMI-GNN demonstrates that session length is a powerful signal for controlling the complexity of multi-interest models. By "softly" merging interests in short sessions and encouraging diversity in longer ones, the model achieves a superior balance between accuracy and diversity. Future work may focus on automating the selection of the session length threshold ($\eta$) and further optimizing the computational efficiency of generating multiple interest vectors.
Summary updated