- The paper introduces a balance-oriented framework employing PER and PRG to enhance personalized video game recommendations.
- It uses statistical normalization and LLM-driven semantic profiling to precisely model user-game interactions and reduce over-smoothing.
- Experimental results on Steam datasets demonstrate significant gains in both accuracy and long-tail diversity compared to leading baselines.
CPGRec+: A Balance-Oriented Framework for Personalized Video Game Recommendation
Introduction and Motivation
The exponential growth of the gaming industry, as exemplified by Steam’s sharply expanding catalog and user engagement, necessitates recommender systems that can simultaneously provide accurate and diverse suggestions. However, the prevailing Graph Neural Network (GNN)-based recommenders predominantly optimize for accuracy, often at the expense of diversity. This focus exacerbates the filter bubble effect, overexposes popular (head) games, and marginalizes long-tail content, ultimately hindering user discovery and market efficiency.
The predecessor to this work, CPGRec, advanced recommenders by constructing category- and popularity-aware game graphs that mediate a trade-off between accuracy (via category coherence) and diversity (via enhanced connectivity and long-tail propagation). Despite these gains, CPGRec's granularity and personalization remained limited due to two factors: it assumed all user-game interactions were equally informative and failed to adequately leverage the semantic reasoning capacities of LLMs. These deficiencies manifest as over-smoothing of player/game representations and insufficient modeling of the variance in user tastes versus global popularity trends.
Framework Overview and Methodological Innovations
CPGRec+ addresses these core limitations through two novel components, culminating in a framework that operationalizes a quantifiable accuracy-diversity balance and advances interaction-level personalization.
PER refines the bipartite player-game graph by assigning signed, information-theoretic edge weights that differentiate not only between significance and insignificance of historical interactions but also between interest and disinterest. The mechanism consists of:
- Fisher Distribution-based Sign Decision: Dwelling time (personal interest) and average rating (global interest) are statistically normalized (Box-Cox + Z-score) and compared per interaction. The Fisher statistic is used to robustly detect significant outlier interactions, distinguishing marked personal preference (interest, disinterest, or neutral).
- Information Content-based Magnitude: The importance of each interaction is further modulated by its information content (negative log-probability in the joint normalized distribution), thereby privileging rare, salient deviations.
This procedure decreases over-smoothing on the bipartite graph, ensuring that edges with little personalized signal do not unduly homogenize node representations via GCN propagation.
PRG exploits LLMs (notably Qwen2.5) to encode deep, context-sensitive game and player profiles, moving beyond surface-level categorical or collaborative signals:
- Rating-Informed Game Descriptions: LLMs are explicitly prompted with salient game metadata and average ratings to generate semantically rich descriptions emphasizing global appeal.
- Preference-Informed Player Descriptions: For each player, LLMs compare personal interaction statistics with global ratings across their game histories, generating unique player embeddings aligned to individual, not population-level, preference trajectories.
These natural language profiles are mapped (e.g., via M3-Embedding and aligned by MLPs) into the representation space and fused with structural graph embeddings for downstream recommendation.
Full Framework Integration
CPGRec+ synthesizes the original trio of CPGRec modules (strict category-based SGC, connectivity-enhanced CNA, popularity-guided PENR) with PER and PRG in a flexible edge/representation reweighting schema. The comprehensive architecture enables tight control over the accuracy-diversity spectrum and delivers robust, long-tail-aware, personalized recommendations.
Experimental Evaluation
Datasets and Metrics
Two real Steam datasets (Steam I, Steam II) are used, incorporating player-game interactions, category/developer/publisher metadata, dwelling time, and ratings. Evaluation covers accuracy (Recall, NDCG, Precision, Hit Rate at K), individual and global diversity (Coverage, Entropy), and long-tail-specific metrics (Tail Coverage, Tail at K).
Empirical Results
- Numerical Superiority: CPGRec+ consistently outperforms state-of-the-art baselines in both accuracy and diversity. In accuracy-focused, diversity-focused, and trade-off (balance) setups, CPGRec+ achieves the best or runner-up scores across all metrics. Gains on long-tail metrics are significant, marking an improved exposure of niche games.
- PER/PRG Ablations: Removal of PER or PRG substantially degrades accuracy and reduces diversity (especially Tail Coverage), confirming their necessity for both individualized modeling and global diversification.
- Balance and Flexibility: CPGRec+ demonstrates a capacity to smoothly interpolate between accuracy and diversity optima, surpassing methods like EXPLORE or DGRec, which often sacrifice one aspect for the other.
- Complexity: The additional computational load of PER and PRG is minimal at inference due to precomputation; CPGRec+ matches standard GNN-based recommenders in scalability and deployability.
Theoretical and Practical Implications
CPGRec+ advances the theoretical modeling of accuracy-diversity trade-offs in recommender systems, particularly by explicitly incorporating preference variance and user-interest dissonance at the interaction level. The work strengthens the case for hybridizing graph-based representations with LLM-generated semantic embeddings, which are shown to be both computationally viable and functionally superior.
On a practical level, the design facilitates broad exposure to long-tail content, which benefits players (richer discovery), developers (increased niche game visibility), and platforms (enhanced user retention and monetization). The explicit modularity in its architecture allows for targeted enhancements, such as swapping in more advanced LLMs or integrating temporal adaptation for evolving preferences.
Future lines of research include dynamic graph updates, enhanced uncertainty quantification for LLM-generated profiles [xfzhou], and multi-modal fusion (e.g., integrating visual/audio cues from game media) to further tailor recommendations.
Conclusion
CPGRec+ represents a significant advance in balance-oriented game recommender systems by rectifying core limitations of GNN over-smoothing and interaction-level modeling. Its integration of statistical edge reweighting and LLM-driven representation learning yields state-of-the-art performance in both accuracy and diversity, especially for long-tail games. The framework's generality and computational tractability make it a promising foundation for adaptive, user-centric recommendation in large and dynamic content ecosystems.
Reference:
"CPGRec+: A Balance-oriented Framework for Personalized Video Game Recommendations" (2604.14586)
