Abstract
Multimodal recommendation systems often struggle in e-commerce settings where user behavior is noisy and not solely driven by item content. We present STAIR (STepwise grAph con-volution), a framework that partitions embedding dimensions to preserve both collaborative and multimodal signals. By introducing Forward and Backward Stepwise Convolutions, STAIR prevents "modality erasure" and "forgetting," achieving state-of-the-art accuracy with significantly lower computational overhead than existing models.
The Problem: Modality Erasure and Forgetting
In multimodal recommendation, models use both user-item interactions (collaborative filtering) and item content like images and text. However, e-commerce behavior is often "noisy"; a user might purchase a product for its utility rather than its aesthetic. Existing Graph Convolutional Network (GCN) approaches suffer from two major flaws in this context:
- Modality Erasure: As GCN layers are stacked to capture collaborative signals, the specific multimodal features (visual/textual) are washed out almost immediately.
- Modality Forgetting: During training, the collaborative signal dominates, causing the model to "forget" the initial multimodal information provided at initialization.
Methodology
STAIR addresses these issues by treating embedding dimensions as specialized storage for different information types. The framework relies on two core mechanisms:
- Forward Stepwise Convolution (FSC): FSC assigns varying weights across embedding dimensions. Early dimensions are weighted to favor deep graph convolutions (capturing collaborative filtering), while later dimensions are preserved to retain raw multimodal features. This creates a "soft transition" within a single embedding vector.
- Backward Stepwise Convolution (BSC): To prevent forgetting, BSC uses a k-Nearest Neighbors (kNN) similarity graph based on raw features. This graph "anchors" the gradient updates, ensuring that items that are multimodally similar receive consistent embedding updates.
Additionally, the model starts with item embeddings initialized via a whitening technique (SVD) on raw multimodal features, rather than starting from random noise.
Key Insights: Quantifying Uncertainty
A key motivation for STAIR is the inherent noise in e-commerce data. The authors used Shannon entropy to prove that multimodal features are sometimes only slightly better than random noise at predicting user behavior.
As shown in the figure, collaborative signals (MF-BPR) have much lower uncertainty than visual or textual features. This justifies STAIR's "complementary" approach, where content is used to enhance rather than replace collaborative data.
Key Results and Efficiency
Evaluated on large-scale Amazon datasets (Baby, Sports, Electronics), STAIR outperformed state-of-the-art models like FREEDOM and LATTICE by 2% to 6% in Recall and NDCG.
Beyond accuracy, STAIR's efficiency is its standout feature. It achieves these results with 1/2 to 1/100 of the computational and memory costs of competing methods. For instance, on the Electronics dataset, STAIR required only 1.7 GB of VRAM, whereas a similar model like MMGCN required 8.5 GB. Its training speed is nearly identical to LightGCN, making it highly scalable for real-world production.
Conclusion
STAIR demonstrates that careful manipulation of the embedding space can solve the fundamental trade-offs in multimodal recommendation. By preventing erasure and forgetting through stepwise convolutions, it provides a scalable, high-performance solution for real-world e-commerce platforms.
Summary updated