A contrastive learning framework for Transformer-based collaborative filtering

doi:10.1016/j.eswa.2026.133508

A contrastive learning framework for Transformer-based collaborative filtering

2026-07-01

1

PRE

AI

DOI：10.1016/j.eswa.2026.133508

Original

Original request for help

Save

Abstract

En 中文

In various machine learning tasks, Transformer-based models have attracted significant attention and achieved notable success. To apply Transformers to Collaborative Filtering (CF), we designed a contrastive learning framework called Transformer Collaborative Filtering (TransCF), in which the Transformer serves as a backbone to extract user preferences from global collaborative signals. During training, instead of traditional user-item pairs, we utilize the entire set of items interacted with by each user to capture high-order item-to-item relationships. Additionally, we integrate a Siamese contrastive learning objective to ensure consistent and coherent user representations across different latent spaces, which significantly enhances performance in sparse and cold-start scenarios. Extensive experiments across diverse domains (including E-commerce, Movie, and News) demonstrate that TransCF achieves state-of-the-art results with superior prediction quality. Furthermore, we provide a scalability analysis confirming O(n) efficiency for large-scale systems and utilize attention-based visualizations to offer intuitive insights into user preferences.

Keywords:

Contrastive learning

Collaborative filtering

Recommendation

Transformer

A contrastive learning framework for Transformer-based collaborative filtering

Abstract

Journal

Organization

Cited Papers