Abstract
In recommender systems, the cold-start issue is challenging due to the lack of interactions between users or items. Such an issue can be alleviated via data-level and model-level strategies. Traditional data-level methods employ auxiliary information like feature information to enhance the learning of user and item embeddings. Recently, Heterogeneous Information Networks (HINs) have been incorporated into the recommender system as they provide more fruitful auxiliary information and meaningful semantics. However, these models are unable to capture the structural and semantic information comprehensively and neglect the unlabeled information of HINs during training. Model-level methods propose to apply the meta-learning framework which naturally fits into the cold-start issue, as it learns the prior knowledge from similar tasks and adapts to new tasks quickly with few labeled samples. Therefore, we propose a contrastive meta-learning framework on HINs named CM-HIN, which addresses the cold-start issue at both data level and model level. In particular, we explore meta-path and network schema views to describe the higher-order and local structural information of HINs. Within meta-path and network schema views, contrastive learning is adopted to mine the unlabeled information of HINs and incorporate these two views. Extensive experiments on three benchmark datasets demonstrate that CM-HIN outperforms all state-of-the-art baselines in three cold-start scenarios.
Introduction
The "cold-start" problem remains a significant hurdle for collaborative filtering-based recommender systems. When new users or items enter a system without a history of interactions, traditional models lack the necessary signal to generate accurate recommendations. To address this, researchers typically employ two strategies: data-level approaches, which leverage auxiliary information, and model-level approaches, such as meta-learning, which aim to generalize from minimal data.
Heterogeneous Information Networks (HINs) have emerged as a powerful data-level tool, representing diverse entities and their multi-faceted relationships as a graph. However, existing HIN-based models often struggle to fully capture local structural nuances or utilize the vast amount of unlabeled data present in these networks.
The CM-HIN Framework
This paper introduces CM-HIN, a framework that integrates meta-learning with contrastive learning on HINs to tackle cold-start scenarios comprehensively. The architecture is built upon three primary components:
1. Meta-Learning Backbone
CM-HIN utilizes a Model-Agnostic Meta-Learning (MAML) approach. By treating the recommendation for a specific user as an individual task, the model learns a global parameter initialization. This allows the system to adapt rapidly to a new user's preferences with only a few interactions.
2. Multi-View Map Encoder
The core innovation lies in how CM-HIN encodes the HIN structure. It processes the network through two distinct lenses:
- Meta-path View: This view captures higher-order semantics, such as "User-Item-Director-Item" (UIDI) paths. It employs contrastive learning to align neighborhood-based perception with specific semantic knowledge.
- Network Schema View: This view focuses on the "blueprint" or local structural information (e.g., direct User-Item or Item-Actor links). To ensure robustness, the model uses a "map drop nodes" technique, forcing the encoder to maintain consistent representations even when certain node types are masked.
3. Contrastive Integration
To bridge these views, CM-HIN performs contrastive learning between the meta-path and network schema representations. This cross-view alignment ensures that the resulting user and item embeddings are enriched by both broad semantic context and immediate local connectivity.
Experimental Evaluation
The model was rigorously tested on three benchmark datasets: DBook (Douban), MovieLens, and Yelp. The researchers evaluated performance across three cold-start scenarios: User Cold-start (UC), Item Cold-start (IC), and User-Item Cold-start (UIC).
Key findings include:
- Superior Accuracy: CM-HIN consistently outperformed state-of-the-art baselines like MeLU and MetaHIN. For instance, in the UIC scenario on the DBook dataset, CM-HIN achieved an NDCG@5 of 0.9114.
- Structural Importance: Ablation studies confirmed that both the meta-path and network schema views are essential; removing either led to a measurable decline in recommendation quality.
- Efficiency: Despite its multi-view architecture, CM-HIN remains computationally efficient, with training times on MovieLens (34.4s per epoch) beating several less accurate meta-learning alternatives.
Conclusion
CM-HIN demonstrates that combining the rapid adaptation of meta-learning with the rich, multi-layered structural information of HINs provides a robust solution to the cold-start problem. By leveraging contrastive learning to mine unlabeled data, the model effectively captures both local and global semantics. Future work will explore dynamic HINs to incorporate temporal user behavior and more sophisticated "hard negative" mining to further refine contrastive representations.
Summary updated