【公开学术报告】Crossing Boundaries: Recommending Serendipity with Cross-Domain Models

Crossing Boundaries: Recommending Serendipity with Cross-Domain Models

Speaker: FU Zhe(The University of North Carolina)

Date & Time: Mon. 24th, March 2025, from 9:30 AM to 11:00 AM (Beijing Time)

Zoom Meeting ID:  87289487669 (Password: 120982)

Join via the Link: https://us02web.zoom.us/j/87289487669.

ABSTRACT

Serendipity means unexpected discoveries that are valuable. Many recommender systems researchers have advocated serendipity to encourage exploratory information discovery in a digital environment. However, modeling serendipity is challenging due to its elusive nature. Collecting large-scale ground truth on serendipity is also challenging due to its subjectivity and sparse occurrences in real life. In this study, we developed a cross-domain recommendation model, called SerenMulti, to model serendipity. It leverages a pre-train and fine-tune mechanism to decompose the elusive serendipity into more tangible notions during the pre-training stage. It also utilizes a cross-domain fine-tuning layer to leverage the knowledge beyond one domain to mitigate the inherent data sparsity problem. For collecting ground truth data, we made use of the recent advances in Large Language Models (LLMs) by instruction tuning two LLMs for labeling whether a piece of text describes a serendipity experience. A total number of 24,992,476 pieces of text in five different domains, such as books, movies, and health, were labeled. The resulting dataset is named SerenMultiLens, the largest ever and multi-domain ground truth data on serendipity. Through a series of experiments, SerenMulti trained on SerenMultiLens outperforms the state-of-the-art baseline models in predicting serendipity. We believe both the model SerenMulti and the dataset SerenMultiLens will empower everyday users with increased chances of bumping into unexpected but valuable discoveries.

Keywords— Cross-domain recommendations, Serendipity, Large Language Models