TL;DR
This paper introduces a session-adaptive propagation framework that enhances session-based recommendations by tailoring global item information to a user's specific intent. By combining a multi-relational knowledge graph with a "query-based" graph neural network, the authors achieved a 10-20% performance boost on standard benchmarks and a 2% lift in a live industrial deployment at The Home Depot.
What Problem Does This Paper Address?
Session-based recommender systems (SBRSs) aim to predict a user's next move based on a short sequence of interactions (a "session"). While modern models like Transformers are excellent at capturing local patterns within a single session, they often struggle to incorporate "global" knowledge—how items relate to each other across millions of other sessions.
Existing hybrid models that use Graph Neural Networks (GNNs) to bring in this global data face two major hurdles:
- Context Blindness: They often aggregate neighbor information "blindly." For example, if a user clicks on a "flower," a standard GNN might pull in data about both "garden loppers" and "kitchen vases." If the user is currently decorating a garden, the vase data is irrelevant noise.
- Uniform Relationships: Most models treat all interactions (viewing, adding to cart, purchasing) as the same, ignoring the nuanced difference between items that are substitutes (co-viewed) versus complements (co-purchased).
Methods
The authors propose a framework consisting of three main components:
- Multi-Typed Knowledge Graph (KG): Instead of a simple graph, they build a KG with three distinct edge types: co-view (items viewed together), co-ATC (items added to cart together), and co-view-ATC (one viewed, the other added to cart). This captures the difference between "this or that" (substitutes) and "this and that" (complements).
- Session-Adaptive Propagation: This is the core innovation. The system uses a two-stage Transformer approach. The first Transformer generates a "contextual embedding" for the items in the current session. This embedding acts as a query to a Heterogeneous Graph Transformer (HGT). The HGT then selectively pulls information from the global KG that is relevant to the specific session's intent.
- Multi-Task Learning: To further sharpen the model, the authors add a secondary task: predicting the metadata (like category or price) of the next item. This forces the model to learn richer, more descriptive item representations.
Datasets
The model was evaluated on three datasets:
- Diginetica: A standard e-commerce benchmark from the CIKM Cup 2016.
- Yoochoose 1/64: A subset of the RecSys Challenge 2015 dataset.
- The Home Depot (THD): A massive, real-world industrial dataset containing over 3 million sessions. This dataset is particularly rich, featuring seven meta-attributes per item (brand, color, manufacturer, etc.).
Key Insights and Ideas
- Intent-Aware Aggregation: The "aha moment" is treating the session context as a filter for global knowledge. By using the current session to "query" the global graph, the model ensures that the item embeddings are dynamically updated to reflect the user's immediate goal.
- The Power of Meta-Attributes: In industrial settings where data can be sparse (the "cold start" problem), leveraging attributes like "Manufacturer" or "Class" allows the model to transfer knowledge between similar items that haven't appeared together in many sessions.
- Relationship Nuance: The ablation studies showed that different edge types matter differently depending on the dataset. For click-heavy datasets (Diginetica), co-view edges were vital. For purchase-oriented sessions, co-ATC edges provided the most signal.
Key Results
- Significant Gains: Equipping standard backbones (like GRU4Rec or SASRec) with this framework improved performance (NDCG@10 and MRR@10) by 10% to 20%.
- State-of-the-Art: The proposed
M-KGHTmodel outperformed existing state-of-the-art baselines like NARM and SR-GNN across all tested datasets. - Real-World Impact: In a live deployment at The Home Depot, the model achieved a 2% performance boost over an already highly-optimized, well-deployed production model.
- Sparsity Resilience: The model proved particularly effective for sessions with "middle-level" sparsity—where there is enough global information to be useful, but the local session data is not yet overwhelming.
Limitations and Future Work
- Node Homogeneity: Currently, the knowledge graph only contains "item" nodes. The authors suggest that adding "user" nodes could allow for even deeper personalization.
- Temporal Shifts: The model treats all historical sessions with equal weight. Future iterations could incorporate temporal dynamics, giving more importance to recent global transition patterns which might better reflect current trends.
- Fairness and Diversity: The authors noted that the model's performance varies based on session sparsity and suggested that future work should focus on balancing performance across different types of sessions to ensure fairness.
Summary updated