Why not Collaborative Filtering in Dual View? Bridging Sparse and Dense Models

TL;DR

The paper tackles the long-standing issue that dense embedding-based collaborative filtering suffers from an $SNR$ ceiling on tail items due to data sparsity. It introduces SaD, a plug-and-play dual-view framework that unifies sparse interaction structure with dense semantic embeddings through bidirectional alignment: the dense view benefits from sparse structural signals and the sparse view gains from dense semantic guidance. The authors provide a theoretical $SNR$ analysis showing how complementary, weakly correlated views can improve fusion performance, and demonstrate empirically that SaD achieves state-of-the-art results across four benchmarks, with pronounced gains for unpopular items. The approach is architecture-agnostic and can boost a wide range of backbones, highlighting the enduring potential of collaborative filtering when exploited from both sparse and dense perspectives. Overall, SaD offers a principled, generalizable method to synergize structure and semantics in recommender systems, with strong practical implications for mitigating popularity bias and enhancing tail-item recommendations.

Abstract

Collaborative Filtering (CF) remains the cornerstone of modern recommender systems, with dense embedding--based methods dominating current practice. However, these approaches suffer from a critical limitation: our theoretical analysis reveals a fundamental signal-to-noise ratio (SNR) ceiling when modeling unpopular items, where parameter-based dense models experience diminishing SNR under severe data sparsity. To overcome this bottleneck, we propose SaD (Sparse and Dense), a unified framework that integrates the semantic expressiveness of dense embeddings with the structural reliability of sparse interaction patterns. We theoretically show that aligning these dual views yields a strictly superior global SNR. Concretely, SaD introduces a lightweight bidirectional alignment mechanism: the dense view enriches the sparse view by injecting semantic correlations, while the sparse view regularizes the dense model through explicit structural signals. Extensive experiments demonstrate that, under this dual-view alignment, even a simple matrix factorization--style dense model can achieve state-of-the-art performance. Moreover, SaD is plug-and-play and can be seamlessly applied to a wide range of existing recommender models, highlighting the enduring power of collaborative filtering when leveraged from dual perspectives. Further evaluations on real-world benchmarks show that SaD consistently outperforms strong baselines, ranking first on the BarsMatch leaderboard. The code is publicly available at https://github.com/harris26-G/SaD.

Figures (6)

• Figure 1: Items are grouped according to the number of user interactions and divided into three categories, from the least to the most interactions. Owing to the low signal-to-noise ratio (SNR), commonly used dense models perform poorly on unpopular items.
• Figure 2: Overview of the SaD model architecture. While each view—Dense and Sparse—performs its own computations, the two views interact through a cross-view alignment mechanism. The Sparse and Dense Alignment Module allows each view to incorporate complementary information from the other. Finally, the Alignment Projector integrates both views, enabling the model to leverage the strengths of each.
• Figure 3: Performance across item popularity groups. Under the SaD framework, the MF model demonstrates substantial improvements on unpopular (long-tail) items.
• Figure 4: Performance of various $\beta$ factors.
• Figure 5: Sensitivity analysis on Movielens and Yelp with respect to the pseudo-positive ratio and top-$K$ hyperparameters.