Empowering Collaborative Filtering with Principled Adversarial Contrastive Loss (2310.18700v1)

Abstract: Contrastive Learning (CL) has achieved impressive performance in self-supervised learning tasks, showing superior generalization ability. Inspired by the success, adopting CL into collaborative filtering (CF) is prevailing in semi-supervised top-K recommendations. The basic idea is to routinely conduct heuristic-based data augmentation and apply contrastive losses (e.g., InfoNCE) on the augmented views. Yet, some CF-tailored challenges make this adoption suboptimal, such as the issue of out-of-distribution, the risk of false negatives, and the nature of top-K evaluation. They necessitate the CL-based CF scheme to focus more on mining hard negatives and distinguishing false negatives from the vast unlabeled user-item interactions, for informative contrast signals. Worse still, there is limited understanding of contrastive loss in CF methods, especially w.r.t. its generalization ability. To bridge the gap, we delve into the reasons underpinning the success of contrastive loss in CF, and propose a principled Adversarial InfoNCE loss (AdvInfoNCE), which is a variant of InfoNCE, specially tailored for CF methods. AdvInfoNCE adaptively explores and assigns hardness to each negative instance in an adversarial fashion and further utilizes a fine-grained hardness-aware ranking criterion to empower the recommender's generalization ability. Training CF models with AdvInfoNCE, we validate the effectiveness of AdvInfoNCE on both synthetic and real-world benchmark datasets, thus showing its generalization ability to mitigate out-of-distribution problems. Given the theoretical guarantees and empirical superiority of AdvInfoNCE over most contrastive loss functions, we advocate its adoption as a standard loss in recommender systems, particularly for the out-of-distribution tasks. Codes are available at https://github.com/LehengTHU/AdvInfoNCE.

• The paper introduces AdvInfoNCE, a novel hardness-aware contrastive loss that dynamically adjusts negative sample difficulty to enhance collaborative filtering performance.
• It leverages adversarial training and DRO frameworks to mitigate exposure bias and false negatives inherent in recommendation systems.
• Experimental results demonstrate significant accuracy improvements in out-of-distribution settings across both synthetic and real-world datasets.

Empowering Collaborative Filtering with Principled Adversarial Contrastive Loss

The paper, "Empowering Collaborative Filtering with Principled Adversarial Contrastive Loss," addresses the application of contrastive learning (CL) in collaborative filtering (CF) systems, proposing an innovative approach called Adversarial InfoNCE (AdvInfoNCE) loss to enhance recommendation performance under implicit feedback scenarios.

Overview and Motivation

Contrastive learning has garnered significant attention due to its efficacy in self-supervised learning tasks, particularly for its capacity to discern meaningful patterns from limited data by maximizing the agreement between positive pairs while distinguishing them from negative pairs. This property is appealing in CF, which often involves semi-supervised top-K recommendation settings focusing on extracting signals from user interactions with items. Existing CL adaptations in CF leverage data augmentation and contrastive losses (like InfoNCE) to improve recommendations, yet face obstacles due to out-of-distribution challenges, false negatives, and the intricacies involved in evaluating top-K recommendations.

The paper identifies two primary limitations in how CL is currently applied to CF:

1. Existing methods disregard CF's inherent characteristics, i.e., exposure bias and false negatives, leading to degraded recommendation quality.
2. There is an inadequate theoretical understanding of how contrastive loss impacts CF models' generalization ability.

Proposed Methodology: AdvInfoNCE

AdvInfoNCE introduces a structured approach to overcome these limitations. It incorporates a hardness-aware ranking criterion that assigns difficulty levels to each negative instance in an adversarial setting, thus refining the model's ability to handle hard and false negatives.

The core innovation is establishing a principled framework that automatically adjusts the hardness of negative samples based on their relative difficulty. This process involves:

• A fine-grained ranking criterion: Enhancing the discrimination of the model by tailoring the ranking criteria to consider the varying difficulty levels of user-item interactions.
• An adversarial training paradigm: Formulating hardness learning as an adversarial problem to enable dynamic learning of interactions' difficulty, empowering the model's generalization capability.

AdvInfoNCE theoretically aligns with Kullback-Leibler (KL) divergence-constrained distributionally robust optimization (DRO) frameworks, emphasizing the formation of robust CF models through effective negative sampling strategies.

Experimental Results and Implications

The research validates AdvInfoNCE across several datasets, both synthetic and real-world, demonstrating its superior performance in out-of-distribution settings compared to state-of-the-art CL-based CF methods. Results reveal substantial improvements in recommendation accuracy, especially under significant distribution shifts, underscoring AdvInfoNCE's robust generalization prowess.

Adopting AdvInfoNCE offers both practical and theoretical benefits:

• Practical: Improves recommendation accuracy by effectively tackling the exposure bias issue and promotes resilience under different dataset conditions by dynamically refining negative samples.
• Theoretical: Enriches the understanding of contrastive loss in CF by linking it to DRO and ensuring a model's stable performance in dynamic real-world environments.

Conclusion and Future Developments

The paper advocates adopting AdvInfoNCE as a standard loss function in recommender systems due to its demonstrated efficacy and theoretical grounding. This approach significantly bridges the gap in the contrastive loss understanding in CF and presents potential avenues for further research into more generalized and robust CF systems. The adaptability of AdvInfoNCE hints at possible extensions into broader domains, signifying its contribution as a substantial reference point in the landscape of collaborative filtering research. The paper's insights into adversarial hardness learning mechanisms promise advancements in CF systems' alignment with inherent recommendation challenges.