MP2: A Momentum Contrast Approach for Recommendation with Pointwise and Pairwise Learning

Menghan Wang, Yuchen Guo, Zhenqi Zhao, Guangzheng Hu, Yuming Shen, Mingming Gong, Philip H. S. Torr

TL;DR

This paper introduces MP2, a novel recommendation framework that tackles the "annotation bias" inherent in implicit feedback data. By combining pointwise and pairwise learning within a momentum contrast framework, MP2 achieves more robust and accurate user preference modeling, outperforming existing state-of-the-art methods.

What Problem Does This Paper Address?

Modern recommendation systems heavily rely on implicit feedback (like clicks or purchases) to understand user preferences. However, this data often comes with "annotation bias," where a single item might receive conflicting implicit signals for the same user across different interactions. For instance, a user might click on item B in one context but not in another, leading to contradictory binary labels (0 or 1). This ambiguity creates conflicting training signals for recommendation models, hindering their ability to accurately capture true user preferences. Traditional pointwise learning methods struggle with this, while pairwise learning, which focuses on relative item order, offers a more robust way to handle preference nuances. The paper aims to bridge this gap by effectively combining the strengths of both approaches.

Methods

The core of MP2 is a three-tower network architecture consisting of one user network and two item networks: a "vanilla" item network and a "momentum" item network.

  • Pointwise and Pairwise Learning: The framework integrates both pointwise and pairwise loss functions. Pointwise loss uses the predicted preference score, while pairwise loss focuses on the relative ordering of items.
  • Momentum Update: To address the annotation bias and ensure stable item representations, the "momentum" item network's weights are updated using a momentum mechanism. This means its parameters evolve more slowly and smoothly than the "vanilla" item network's, effectively creating a consistent, slowly-evolving item representation that is less susceptible to label fluctuations.
  • Weighting Label with Discrepancy: The paper introduces a novel way to measure the "annotation bias" by calculating the discrepancy between the representations from the vanilla and momentum item networks. This discrepancy serves as a confidence score for the pointwise labels. A higher discrepancy indicates more uncertainty, leading to a lower weight for that specific pointwise label in the loss calculation. This automatically down-weights potentially noisy or biased labels.

Datasets

The paper evaluates MP2 on four publicly available datasets:

  • MovieLens-100k
  • MovieLens-1M
  • Beauty (from Amazon)
  • Office Products (from Amazon)

These datasets are commonly used for recommendation system research. For pointwise labels, explicit ratings were binarized using a threshold of 3. For pairwise labels, item pairs were sampled under the same user, and their relative order was determined by their scores.

Key Insights and Ideas

  • Annotation bias is a real problem: The paper convincingly demonstrates how conflicting implicit feedback can lead to significant issues in training recommendation models.
  • Pairwise learning is crucial for robustness: The study highlights that pairwise learning, by focusing on relative preferences, inherently avoids the annotation bias problem of binary pointwise labels.
  • Momentum contrast for stable representations: Applying momentum updates to item representations, inspired by self-supervised learning in computer vision, is shown to be effective in creating stable embeddings that are less affected by noisy labels.
  • Discrepancy as a confidence measure: Quantifying the difference between a "fast-evolving" and a "slow-evolving" item representation provides a direct measure of label uncertainty, which can then be used to dynamically adjust the influence of pointwise labels.
  • Synergy of pointwise and pairwise: The combination of pointwise and pairwise learning, enhanced by the momentum contrast mechanism, leads to superior performance, suggesting that these two learning paradigms are complementary.

Key Results

MP2 consistently outperformed several state-of-the-art baselines across all four datasets in terms of HitRate@k and NDCG@k metrics. For example, on MovieLens-100k, MP2 achieved an NDCG@5 of 0.4721, surpassing other methods. The paper also shows that:

  • Jointly learning pointwise and pairwise objectives (as in MP2 and T3) is generally better than using either alone.
  • The momentum update strategy is effective, with performance increasing as the momentum coefficient ($\alpha$) approaches 1 (indicating a smoother update).
  • The discrepancy-based weighting strategy (MP2 joint) is superior to uniform weighting (MP2uniform) or separate weighting (MP2separate), demonstrating the value of dynamically adjusting label weights based on representation uncertainty.

Limitations and Future Work

  • Limitations: The paper focuses on pointwise and pairwise learning, acknowledging that listwise learning is beyond the scope of this work. The effectiveness of the discrepancy weighting strategy is demonstrated on two datasets, but further validation on more diverse datasets could be beneficial.
  • Future Work: The authors propose extending MP2 to incorporate listwise loss functions alongside pointwise and pairwise learning for even more comprehensive recommendation.

Summary updated

Abstract

Binary pointwise labels (aka implicit feedback) are heavily leveraged by deep learning based recommendation algorithms nowadays. In this paper we discuss the limited expressiveness of these labels may fail to accommodate varying degrees of user preference, and thus lead to conflicts during model training, which we call annotation bias. To solve this issue, we find the soft-labeling property of pairwise labels could be utilized to alleviate the bias of pointwise labels. To this end, we propose a momentum contrast framework (MP2) that combines pointwise and pairwise learning for recommendation. MP2 has a three-tower network structure: one user network and two item networks. The two item networks are used for computing pointwise and pairwise loss respectively. To alleviate the influence of the annotation bias, we perform a momentum update to ensure a consistent item representation. Extensive experiments on real-world datasets demonstrate the superiority of our method against state-of-the-art recommendation algorithms.
BibTeX
@inproceedings{wang2022mp,
  author = {Menghan Wang and Yuchen Guo and Zhenqi Zhao and Guangzheng Hu and Yuming Shen and Mingming Gong and Philip H. S. Torr},
  title = {{MP2: A Momentum Contrast Approach for Recommendation with Pointwise and Pairwise Learning}},
  booktitle = {International ACM SIGIR Conference on Research and Development in Information Retrieval},
  year = {2022},
}
Figure