SerenDiff: Generating Serendipity Recommendations through a Diffusion Model

Zhe Fu, Xi Niu

Keywords: serendipity, recommendations, Diffusion Models

TL;DR: In this project, we leveraged a conditional diffusion model as a recommendation model to recommend serendipity.

Abstract: Serendipity means an unexpected but valuable discovery. It has attracted wide attention in recommender systems research in recent years. Due to its elusive and subjective nature, serendipity is difficult to model even with today's advances in machine learning and deep learning techniques. In addition, most existing serendipity models lack interpretability. To address the modeling challenges and the interpretability issues, we propose a serendipity diffusion recommendation model (named SerenDiff), to generate serendipity recommendations leveraging a state-of-the-art generative AI model, the diffusion model. We regarded a user history with a recommender system as an "image", and the serendipity recommendation generation as a recovering process of the corrupted "image". Diffusion models are believed to be creative in the recovering process of a noised image in the sense that they base on but go beyond the original training data, providing room for finding serendipity. Extensive experiments have shown the effectiveness of SerenDiff. We believe SerenDiff will empower everyday users, not only with increased chances of encountering unexpected but relevant discoveries, but also with explanations on the elusive serendipity recommendations.

Primary Area: generative models

Submission Number: 10097