Adaptive preference learning from multi-user feedback

Loading...
Thumbnail Image

URL

Journal Title

Journal ISSN

Volume Title

School of Science | Master's thesis

Department

Mcode

Language

en

Pages

59

Series

Abstract

We consider the problem of multi-user preference learning, where the model learns from multiple human users and uses the learned information to quickly adapt to new users. Previous attempts at developing such a system assumed that users who are connected on platforms like social media tend to like similar products. This assumption is unlikely to hold in practice because users usually know only a small fraction of their connections. Other works imposed a structure on user preferences using shared latent functions. This method is limited by the representation capabilities of the chosen latent functions. In this work, we proposed a novel probabilistic approach to multi-user preference learning by partitioning user preferences into two components: an (optional) shared knowledge between users and a set of individual parameters. We trained a model on each user in the training user base and aggregated these trained models to extract both the shared and individual components simultaneously. The extracted information is stored using a Bayesian Neural Network, which facilitates efficient adaptation to new users. During learning and adaptation, we employed active learning to query items to users for labeling. We tested this framework on both synthetic and real preference datasets, and also applied the method to an existing retrosynthesis scoring system. The results show that our model outperformed even personalized models trained on the users in terms of recommending high-rating items to users.

Description

Supervisor

Kaski, Samuel

Thesis advisor

Guo, Yujia

Other note

Citation