Consensus Learning from Heterogeneous Objectives for One-Class Collaborative Filtering

Seongku Kang, Dongha Lee, Wonbin Kweon, Junyoung Hwang, Hwanjo Yu

Introduction

In One-Class Collaborative Filtering (OCCF), models must learn user preferences from implicit feedback, such as clicks or views, where only positive signals are observed. Researchers have developed numerous learning objectives (loss functions) to address this, including Bayesian Personalized Ranking (BPR), Collaborative Metric Learning (CML), and Binary Cross-Entropy (BCE). However, no single objective consistently outperforms others across all datasets. This paper introduces ConCF (Consensus learning for OCCF), a framework that leverages the complementary strengths of multiple heterogeneous objectives to improve recommendation accuracy.

Methodology

ConCF employs a multi-branch architecture during training to capture diverse facets of user behavior. The framework consists of a shared user-item encoder and multiple auxiliary heads, each optimized with a different learning objective (e.g., BPR, CML, BCE, MSE, and Multinomial).

The ConCF architecture (b) compared to a standard single-objective model (a). The framework uses multiple heads and a consensus mechanism during training.

Ranking-Based Consensus

Because different loss functions produce scores in varying ranges, ConCF converts these scores into relative rankings to create a "universal language" for knowledge exchange. A central Consensus module aggregates these rankings. To ensure the consensus is reliable, the model utilizes temporal consistency: it identifies stable predictions across training epochs as more likely to be correct, weighting the consensus toward these reliable signals.

Collaborative Evolution and Inference

During training, each head is pulled toward the collective consensus while simultaneously satisfying its specific objective. To balance these potentially conflicting gradients, the framework uses dynamic gradient normalization (GradNorm). Crucially, the auxiliary heads are discarded after training, resulting in a standard single-head model that maintains inference efficiency while benefiting from multi-objective knowledge.

Empirical Motivation and Results

The authors demonstrate that different objectives are indeed complementary using the Pairwise Exclusive-hit Ratio (PER). Even when one objective is statistically superior, other objectives often correctly predict items that the "better" model misses.

Performance comparison (R@50) and complementarity heatmaps across three datasets. The heatmaps show that different objectives capture unique user preferences.

Key findings from evaluations on CiteULike, Ciao, and Foursquare include:

  • Significant Accuracy Gains: ConCF consistently outperformed standalone objectives, improving Recall@20 by approximately 25% on the CiteULike dataset compared to the best single-objective baseline.
  • Superiority over Ensembles: The framework outperformed traditional model ensembles, suggesting that the interaction between objectives during training is more effective than post-hoc combination.
  • Robustness for Sparse Users: The consensus approach proved particularly beneficial for "cold" users with few interactions, where the shared knowledge helps fill data gaps.

Conclusion

ConCF provides a robust framework for improving OCCF by harmonizing heterogeneous learning objectives. By identifying stable, consistent predictions across multiple heads, the model learns a more comprehensive representation of user preferences. Future work will explore extending this consensus mechanism to more complex architectures like Graph Neural Networks and Transformers.

Summary updated

Abstract

Over the past decades, for One-Class Collaborative Filtering (OCCF), many learning objectives have been researched based on a variety of underlying probabilistic models. From our analysis, we observe that models trained with different OCCF objectives capture distinct aspects of user-item relationships, which in turn produces complementary recommendations. This paper proposes a novel OCCF framework, named ConCF, that exploits the complementarity from heterogeneous objectives throughout the training process, generating a more generalizable model. ConCF constructs a multi-branch variant of a given target model by adding auxiliary heads, each of which is trained with heterogeneous objectives. Then, it generates consensus by consolidating the various views from the heads, and guides the heads based on the consensus. The heads are collaboratively evolved based on their complementarity throughout the training, which again results in generating more accurate consensus iteratively. After training, we convert the multi-branch architecture back to the original target model by removing the auxiliary heads, thus there is no extra inference cost for the deployment. Our extensive experiments on real-world datasets demonstrate that ConCF significantly improves the generalization of the model by exploiting the complementarity from heterogeneous objectives.
BibTeX
@inproceedings{kang2022consensus,
  author = {Seongku Kang and Dongha Lee and Wonbin Kweon and Junyoung Hwang and Hwanjo Yu},
  title = {{Consensus Learning from Heterogeneous Objectives for One-Class Collaborative Filtering}},
  booktitle = {The Web Conference},
  year = {2022},
  doi = {10.1145/3485447.3512070},
}
Figure