Boosting Factorization Machines via Saliency-Guided Mixup.

Area of Science:

• Machine Learning
• Recommender Systems
• Data Mining

Background:

• Factorization Machines (FMs) are effective for recommender systems, especially with sparse data.
• Existing FMs struggle to model direct interactions of non-interactive features, limiting their expressive power.
• Current data augmentation for FMs often requires significant expertise and resources.

Purpose of the Study:

• To propose novel methods, MixFM and Saliency-guided Mixup Factorization Machine (SMFM), to enhance the performance of Factorization Machines.
• To enable FMs to learn direct interactions between non-interactive features in sparse data.
• To improve the generalization ability and expressive power of Factorization Machines.

Main Methods:

• MixFM: Generates auxiliary training data via convex combinations of existing data, inspired by Mixup.
• SMFM: Employs saliency guidance to create more informative augmented data, addressing potential redundancy from MixFM.
• Theoretical analysis to demonstrate minimization of generalization error bounds.

Main Results:

• Both MixFM and SMFM significantly improve the performance of Factorization Machines across seven diverse datasets.
• The proposed methods effectively learn direct interactions for previously non-interactive features.
• Experimental results confirm the superiority of MixFM and SMFM over baseline approaches.

Conclusions:

• MixFM and SMFM offer effective strategies to enhance Factorization Machines, particularly for sparse data scenarios.
• The augmentation techniques improve model expressiveness and generalization by enabling direct feature interactions.
• The study highlights the benefit of carefully generated 'poisoned' or mixed data for improving FM variants.