A knowledge tracing-like approach to modeling dynamic user preferences

Individual preferences change over time, requiring recommendation systems that adapt and provide personalized suggestions. This paper introduces a novel approach called Preference Tracing, inspired by knowledge tracing from the educational domain. Knowledge tracing estimates a student’s knowledge state from interactions with question-response pairs and knowledge components, which are essential for solving given exercises. Based on the estimated knowledge state, the model predicts the probability of correctly answering subsequent exercises. Similarly, Preference Tracing estimates a user’s preference state from rating histories, including movie-rating pairs and a movie component. Movie plots were crawled from Wikipedia, IMDb, and Letterboxd, and then latent Dirichlet allocation (LDA) was applied to define each film’s top-weighted topic as a movie component. Based on that, Preference Tracing can track users’ changing preferences and predict whether a user would like a given movie. Our main contribution demonstrates that Preference Tracing delivers hyper-personalized recommendations by adapting to changing individual preferences. Experimental results on MovieLens 1M show that Preference Tracing outperforms traditional baseline models and effectively captures dynamic changes.