Rethinking Collaborative Metric Learning: Toward an Efficient Alternative Without Negative Sampling.

Area of Science:

• Recommendation Systems
• Machine Learning
• Data Mining

Background:

• Collaborative Metric Learning (CML) is a popular recommendation system paradigm known for its simplicity and effectiveness.
• Existing CML methods often rely on negative sampling to manage computational costs associated with pairwise comparisons.
• Negative sampling, while efficient, can introduce biases in estimating the model's generalization error.

Purpose of the Study:

• To theoretically analyze the impact of negative sampling on CML generalization error.
• To propose a novel CML approach that avoids negative sampling and its associated biases.
• To demonstrate the effectiveness of the proposed method through empirical evaluation.

Main Methods:

• Theoretical analysis of generalization error bounds in sampling-based CML.
• Quantification of sampling bias using per-user Total Variance (TV).
• Development and implementation of Sampling-Free Collaborative Metric Learning (SFCML).

Main Results:

• Negative sampling introduces a bias term in the generalization error bound, dependent on Total Variance (TV).
• Optimizing sampling-based CML does not guarantee small generalization error, even with large datasets.
• The proposed SFCML method effectively removes sampling bias.
• Experiments on seven benchmark datasets show SFCML outperforms existing methods.

Conclusions:

• Negative sampling in CML leads to a fundamental bias in generalization error estimation.
• SFCML offers a practical and theoretically sound alternative to sampling-based CML.
• The proposed SFCML algorithm demonstrates superior performance in recommendation systems.