Abstract
Basket recommendation (BR) aims to predict items a user will likely add to their current shopping session. Traditional models often assume a single, unified intent per basket, which fails to capture the complexity of real-world shopping trips. We present the Multi-Intent Translation Graph Neural Network (MITGNN), a framework that disentangles latent intents as geometric translations within a heterogeneous graph. MITGNN demonstrates significant performance improvements in both transductive and inductive settings across large-scale datasets.
Introduction: The Challenge of Multi-Intent Baskets
Basket Recommendation (BR) is a critical task in e-commerce, involving the prediction of items to complete a user's current shopping cart. Most existing methods operate under the assumption that items within a single basket are semantically related through a single, unified intent—for instance, purchasing bread and eggs for a "breakfast" goal.
However, real-world consumer behavior is often multi-faceted. A single shopping trip might combine disparate needs, such as "breakfast" items (milk, cereal) alongside "cleaning" supplies (detergent, sponges). Forcing these unrelated items into a single semantic embedding "spoils" the representation space, leading to inaccurate recommendations. These intents are typically latent (unobserved) and can be highly correlated.
Methodology: Multi-Intent Translation Graph Neural Network (MITGNN)
To address the limitations of single-intent models, we propose MITGNN, which models a heterogeneous "basket graph" consisting of Users ($u$), Baskets ($b$), and Items ($i$).
Core Components
- Multi-Intent Generator: Drawing inspiration from Knowledge Graph embedding techniques like TransE, this module treats a basket as a "head entity" and generates $T$ distinct "tail entities" (intents) via translation vectors. An intent embedding is defined as the basket embedding plus a learned relation vector: $h = e + o$.
- Multi-Intent Aggregator: Recognizing that different intents vary in relevance to specific users or items, the model employs three attention mechanisms:
- Basket-guided: Aggregates intents to form the overall basket representation.
- User-guided: Aligns intents with a user's long-term historical preferences.
- Item-guided: Models the relationship between specific items and various intents.
- Graph Neural Network (GNN) Layers: Multiple GNN layers propagate these multi-intent signals, capturing high-order interactions across the graph. The authors found that 3 layers of GNN depth were optimal for balancing signal propagation and over-smoothing.
Transductive and Inductive Settings
MITGNN is designed to handle two primary scenarios:
- Transductive BR: Completing a basket that was partially observed during the training phase.
- Inductive BR: Recommending items for an entirely new basket not seen during training.
Experimental Evaluation
The model was evaluated on two large-scale real-world datasets:
- Instacart: A public grocery dataset with ~3 million records (~65k nodes, 1.2M interactions).
- Walmart Grocery: A proprietary dataset from Walmart.com (~85k nodes, 1.2M interactions). The Walmart dataset is significantly sparser (0.0168% density), providing a rigorous test for model robustness.
Key Insights and Results
- Geometric Disentanglement: Treating intents as geometric translations prevents unrelated items from pulling a single embedding in conflicting directions. This preserves the semantic integrity of item embeddings.
- Performance Gains: MITGNN consistently outperformed state-of-the-art baselines (e.g., BPR-MF, Triple2vec, NGCF). On Instacart, it achieved a 9-12% improvement in Recall and NDCG. On the sparser Walmart dataset, improvements exceeded 17%.
- Optimal Intent Count: Empirical testing showed that the optimal number of intents ($T$) is typically 3 or 4, confirming that multi-intent modeling is superior to the standard single-intent approach ($T=1$).
- Inductive Robustness: MITGNN effectively infers embeddings for new baskets by attentively aggregating the embeddings of items already present in the cart.
Conclusion and Future Work
MITGNN provides a robust framework for disentangling latent shopping intents, leading to superior basket recommendations. While the model introduces additional computational complexity due to multiple intent heads, the accuracy gains—especially in sparse data environments—are substantial. Future work will explore integrating temporal dynamics and more complex inter-intent relationships into the graph structure.
Summary updated