Rethinking Semantic Collaborative Integration: Why Alignment Is Not Enough

In recent years, large language models (LLMs) have made significant advances in the field of natural language processing and have rapidly become an integral part of modern recommender systems. Traditional recommender systems primarily rely on collaborative filtering techniques, which predict user preferences based on user-item interaction data. However, this approach has limitations in addressing cold-start problems and long-tail items. To overcome these challenges, researchers have begun integrating LLM-derived semantic embeddings with collaborative representations, hoping to enhance recommendation performance through representation alignment. Although this approach has been successful in some cases, it assumes that semantic and collaborative representations can be integrated through simple geometric alignment, overlooking potential structural differences between them.

The core problem addressed in this paper is the reevaluation of the integration method for semantic and collaborative representations. Traditional representation alignment methods assume that the two views encode a shared latent entity and that stronger alignment yields better results. However, this assumption is overly strong and often structurally mismatched with real-world recommendation settings. Semantic representations are shaped by language modeling objectives and content statistics, reflecting explicit attribute coherence and conceptual similarity, while collaborative representations are induced from interaction graphs encoding co-occurrence patterns, exposure mechanisms, popularity dynamics, and collective behavioral regularities. Although these two views overlap, they arise from different generative processes and may encode distinct, view-specific factors.

The core innovations of this paper include proposing a complementarity fusion perspective that treats semantic and collaborative representations as partially shared yet fundamentally heterogeneous views. Specific innovations include:

1. Introducing a shared-plus-private latent structure model, emphasizing the importance of preserving view-specific signals and selective information sharing.

2. Developing complementarity diagnostics to quantify overlap, unique-hit contribution, and theoretical fusion upper bounds, providing a new analytical perspective.

3. Empirically demonstrating the effectiveness of complementarity fusion strategies, showcasing advantages in handling sparse data and long-tail problems.

The methodology of this paper includes the following key steps:

• �� Treating semantic and collaborative representations as partially shared yet fundamentally heterogeneous views, each containing shared and view-specific factors.
• �� Developing complementarity-aware diagnostics to quantify overlap, unique-hit contribution, and theoretical fusion upper bounds.
• �� Conducting empirical analyses on sparse recommendation benchmarks to validate the item-level agreement and complementarity between semantic and collaborative views.
• �� Performing controlled alignment probes to evaluate the ability of low-capacity mappings to capture shared components and recover full collaborative geometry.

The experimental design includes conducting systematic experiments across multiple recommendation settings using the Movies, Books, and Games datasets. Each dataset undergoes standard 5-core filtering, exhibiting extreme sparsity (>99.9%), making them ideal testbeds for evaluating long-tail signal preservation and severe interaction scarcity. The baseline models used in the experiments include the collaborative filtering model LightGCN and the semantic model BGE-M3, representing the collaborative and semantic views, respectively. By comparing different alignment strategies, the effectiveness of complementarity fusion strategies is evaluated.

The experimental results show that semantic and collaborative views have low item-level agreement and substantial oracle fusion gains, indicating strong complementarity. Specifically, on the Movies dataset, the semantic model achieves a Recall@20 of 0.1136, while the collaborative model achieves 0.1100, demonstrating their independent predictive power. Controlled alignment probes show that low-capacity mappings capture only shared components and fail to recover full collaborative geometry, especially under distribution shifts. These findings suggest that alignment should not be treated as the default integration principle.

The complementarity fusion method proposed in this paper has wide-ranging applications in recommender systems. Direct applications include improving recommendation performance for cold-start users and long-tail items, enhancing the robustness and informational diversity of recommender systems. The industry impact lies in better meeting users' personalized needs, increasing user satisfaction and platform user retention.

Despite the theoretical and experimental success of the method proposed in this paper, there are still some limitations. First, complementarity fusion methods may have higher computational complexity than traditional alignment methods, as they require separate handling of shared and private components. Second, in extremely sparse datasets, more complex mechanisms may be needed to ensure model robustness. Additionally, in some cases, complementarity fusion may not fully utilize all available information, particularly when the proportion of shared information between views is high.