Abstract: With the popularity of social media, an increasing number of users are joining group activities on online social platforms. This elicits the requirement of group identification (GI), which is to recommend groups to users. We reveal that users are influenced by both group-level and item-level interests, and these dual-level interests have a collaborative evolution relationship: joining a group expands the user's item interests, further prompting the user to join new groups. Ultimately, the two interests tend to align dynamically. However, existing GI methods fail to fully model this collaborative evolution relationship, ignoring the enhancement of group-level interests on item-level interests, and suffering from false-negative samples when aligning cross-level interests. In order to fully model the collaborative evolution relationship between dual-level user interests, we propose CI4GI, a Collaborative Interest-aware model for Group Identification. Specifically, we design an interest enhancement strategy that identifies additional interests of users from the items interacted with by the groups they have joined as a supplement to item-level interests. In addition, we adopt the distance between interest distributions of two users to optimize the identification of negative samples for a user, mitigating the interference of false-negative samples during cross-level interests alignment. The results of experiments on three real-world datasets demonstrate that CI4GI significantly outperforms state-of-the-art models.
1. Introduction
Group Identification (GI) is the task of recommending online groups to users, a critical function for platforms like Steam, Mafengwo, and Weeplaces. Unlike group recommendation, which suggests items to a collective, GI focuses on individual user preferences for communities. The core challenge in GI is that users join groups based on two intertwined levels of interest: item-level interests (e.g., liking a specific multiplayer shooter game) and group-level interests (e.g., enjoying highly competitive gaming communities).
2. Motivation: Collaborative Evolution
The authors identify two key phenomena in user behavior that define the "collaborative evolution" of interests:
- Mutual Enhancement: Joining a group exposes users to new items, expanding their item-level interests.
- Dynamic Convergence: Over time, a user's item-level and group-level interests tend to align as they find communities that match their specific tastes.
Existing GI models treat these two levels of interest separately, missing their mutual enhancement. Furthermore, when modern models try to align these dual-level interests using contrastive learning, they often accidentally push away the representations of users with similar interests (treating them as "false-negative samples"). This interference degrades the model's ability to learn accurate user profiles.
3. The CI4GI Framework
To solve these issues, the authors propose CI4GI (Collaborative Interest-aware model for Group Identification). The architecture relies on a mix of Hypergraph Convolutional Networks (HyperGCN) and Graph Attention Networks (GATs) to process User-Group, Group-Item, and User-Item relationships.
Key novelties in the framework include:
- Interest Enhancement Strategy: Instead of just looking at what items a user has interacted with, CI4GI treats the groups a user has joined as "context." It pulls in items interacted with by the user's groups to supplement and expand the user's item-level interest representation.
- Distributional Representations: Rather than representing a user's item-level interest as a fixed point (embedding) in space, CI4GI models it as an independent Gaussian distribution. This captures the variance and flexibility of human preferences.
- Dynamic False-Negative Optimization: CI4GI uses contrastive learning to align a user's group-level and item-level interests. To avoid the false-negative problem, the model calculates the Wasserstein distance between users' interest distributions. If the distance is below a certain threshold—meaning the users have highly overlapping interests—that user is dynamically removed from the negative sample pool for that training batch.
4. Experimental Evaluation
The authors evaluated CI4GI on three publicly available real-world datasets: Mafengwo (travel), Weeplaces (location-based social), and Steam (online gaming).
Key Results
- State-of-the-Art Performance: CI4GI outperformed 8 baselines (including LightGCN, ConsRec, and DiRec) across all three datasets. It achieved an average improvement of 13.09% on the NDCG@10 metric compared to the best-performing baselines.
- Cold-Start Resilience: CI4GI demonstrated superior cold-start performance. By leveraging group-item interactions to enhance user profiles, the model made highly accurate predictions even for users who had historically only joined a single group.
- Ablation Success: Removing either the contextual enhancement module or the dynamic false-negative filtering caused significant performance drops, proving both novel components are essential to the model's success.
5. Conclusion
CI4GI advances Group Identification by explicitly modeling the co-evolution of dual-level user interests. By treating user interests as distributions and dynamically filtering false negatives, the model provides a more robust and accurate recommendation signal. Future work may involve scaling these distributional representations to massive, industry-scale social graphs where interaction density is even higher.
Summary updated