Leveraging ChatGPT to empower training-free dataset condensation for content-based recommendation

Abstract

Modern Content-Based Recommendation (CBR) techniques utilize item content to deliver personalized services, effectively mitigating information overload. However, these methods often require resource-intensive training on large datasets. To address this issue, we explore dataset condensation for textual CBR in this paper. Dataset condensation aims to synthesize a compact yet informative dataset, enabling models to achieve performance comparable to those trained on full datasets. Applying existing approaches to CBR presents two key challenges: (1) the difficulty of synthesizing discrete texts and (2) the inability to preserve user-item preference information. To overcome these limitations, we propose TF-DCon, an efficient dataset condensation method for CBR. TF-DCon employs a prompt-evolution module to guide ChatGPT in condensing discrete texts and integrates a clustering-based module to condense user preferences effectively. Extensive experiments conducted on three real-world datasets demonstrate TF-DCon's effectiveness. Notably, we are able to approximate up to 97% of the original performance while reducing the dataset size by 95% (i.e., dataset MIND). We have released our code and data for other researchers to reproduce our results.