Abstract
While the mining of modalities is the focus of most multimodal recommendation methods, we believe that how to fully utilize both collaborative and multimodal information is pivotal in e-commerce scenarios where user behaviors are rarely determined entirely by multimodal features. We identify two key challenges: modality erasure and modality forgetting. We propose STAIR, which employs a novel Stepwise Graph Convolution to enable a co-existence of collaborative and multimodal information. STAIR achieves state-of-the-art performance with minimal computational and memory costs.
Introduction: The Challenge of Multimodal Recommendation
Modern e-commerce recommendation systems aim to integrate collaborative filtering (CF) signals—derived from user-item interaction history—with multimodal information, such as product images and textual descriptions. However, standard Graph Convolutional Networks (GCNs) often fail to balance these distinct data types effectively.
The authors identify two primary hurdles in current GCN-based recommenders:
- Modality Erasure: Standard graph convolutions "smooth" features across the graph. While beneficial for collaborative filtering, this process rapidly washes out the distinct, fine-grained multimodal features after just a few layers.
- Modality Forgetting: As models optimize for recommendation loss (e.g., BPR loss), they tend to prioritize interaction patterns so heavily that the initial multimodal characteristics are "forgotten" during training.
The Uncertainty Paradox in E-Commerce
A critical insight of this work is the "Uncertainty Paradox." By measuring Shannon entropy to quantify user behavior uncertainty, the researchers found that in e-commerce, textual and visual features are often significantly less predictive than collaborative signals, sometimes performing only slightly better than random noise.
This finding suggests that in e-commerce, multimodal data should serve as a complement to collaborative filtering rather than a primary driver, necessitating a specialized architecture that can preserve both signals without one overwhelming the other.
The STAIR Methodology
To address these challenges, the researchers proposed STAIR (STepwise grAph convolution). STAIR employs a "divide and conquer" strategy by partitioning the dimensions of item embeddings.
Forward Stepwise Convolution (FSC)
Instead of applying uniform convolution across all embedding dimensions, STAIR assigns varying "layer weights." Some dimensions undergo deep convolution to capture rich collaborative signals, while others remain "shallower," preserving raw multimodal features. This allows the model to be deep and shallow simultaneously.
Backward Stepwise Convolution (BSC)
To prevent modality forgetting, STAIR implements a constrained update mechanism. By building a k-Nearest Neighbor (kNN) graph based on original features, the model ensures that items originally similar in appearance or description maintain that similarity in the final embedding space during backpropagation.
Experimental Results and Efficiency
STAIR was evaluated on three Amazon datasets: Baby, Sports, and Electronics. It consistently outperformed previous state-of-the-art models (like FREEDOM) by 2% to 6% in Recall@20 and NDCG@20.
Beyond accuracy, STAIR offers massive efficiency gains. On the large Electronics dataset, STAIR required only 1.7 GB of GPU memory, a fraction of the 8.5 GB required by MMGCN. Furthermore, its training speed is significantly higher than competitors that rely on complex data augmentation or contrastive learning.
Conclusion
STAIR demonstrates that effective multimodal recommendation in e-commerce requires careful manipulation of how information flows through embedding dimensions. By solving modality erasure and forgetting through stepwise convolution, STAIR provides a lightweight, high-performance solution for real-world recommendation tasks.
Summary updated