Introduction
Modern recommendation systems, such as those powering TikTok or Amazon, primarily rely on Graph Neural Networks (GNNs) to model user-item interactions. Typically, these interactions are represented as bipartite graphs where an edge signifies a simple interaction (e.g., User A liked Item B). However, this standard approach faces two significant challenges:
- Contextual Limitations: Standard graphs struggle to capture "higher-order" relationships or indirect interactions that span multiple nodes. To model these deep connections, GNNs require numerous layers, which often leads to "over-smoothing"—a phenomenon where node representations become indistinguishable.
- Multimodal Complexity: Micro-videos integrate visual, auditory, and textual modalities. Existing models often fail to capture the intricate, dataset-wide dependencies between these diverse data types.
To address these issues, we present the Self-Supervised Hypergraph Learning (SHL) framework, which leverages hypergraphs to capture complex global relationships and employs self-supervised learning for robust multimodal representation.
Methodology
The SHL framework processes data through a multi-layered architecture designed to balance local interactions with global context.
Multimodal Embedding and Local Encoding
Raw video, audio, and text features are first extracted using pre-trained models such as ResNet50, VGGish, and Doc2Vec. A Graph Attention Network (GAT) then performs local encoding, focusing on direct user-item interactions to capture immediate neighborhood features.
Global Encoding (The Hypergraph Core)
Unlike standard edges, hyperedges can connect an arbitrary number of nodes simultaneously. This structure allows the model to capture global patterns and higher-order connectivity without the need for deep GNN architectures. To maintain efficiency and scalability, the framework factorizes the hypergraph incidence matrix into low-rank matrices, preventing parameter explosion.
Self-Supervised Alignment
SHL utilizes an InfoNCE loss to align local and global views. The global hypergraph representation serves as a supervisory signal for the local representation, effectively filtering out interaction noise (e.g., accidental clicks) and focusing on salient features.
Key Insights and Ideas
A core contribution of SHL is the conceptualization of hyperedges as "information highways." These highways allow signals to traverse between distant but semantically related nodes much faster than traditional edge-hopping in bipartite graphs. This mechanism effectively resolves the "long-distance" problem in recommendation systems while avoiding the pitfalls of deep GNNs.
Experimental Evaluation
The researchers evaluated SHL on two primary datasets: TikTok (micro-videos with text descriptions) and MovieLens-10M (augmented with YouTube trailers). Both datasets were cleaned to ensure multimodal completeness and have been made publicly available.
Performance and Robustness
SHL consistently outperformed state-of-the-art baselines, including NGCF, LightGCN, and MMGCL. Key results include:
- Significant Gains: SHL achieved performance improvements of up to 5.6% on TikTok and 5.4% on MovieLens across metrics such as Precision@10 and NDCG@10.
- Ablation Studies: Removing either the hypergraph component or the self-supervised learning (SSL) module resulted in significant performance drops, confirming their critical roles in the model's success.
- Resistance to Over-smoothing: Using the Mean Average Distance (MAD) metric, SHL demonstrated the ability to maintain distinct, high-quality node representations even as hidden dimensions increased, whereas traditional models tended to degrade.
Conclusion
The SHL framework demonstrates that combining local graph attention with global hypergraph structures and self-supervised alignment significantly enhances multimodal recommendation. While the model is robust, future work will focus on more advanced denoising techniques for highly noisy datasets and adapting the framework for real-time, dynamic graph structures.
Summary updated