From knowledge tracing to preference tracing: Capturing dynamic user preferences for personalized recommendation

Individual preferences change over time. While sequential recommenders have gained attention for accommodating changing user preferences, they have struggled to identify users’ preferences at a granular, component-wise level. This paper introduces a novel approach called preference tracing, inspired by the concept of knowledge tracing, originally developed in the educational domain. Knowledge tracing dynamically estimates a student's knowledge state through interactions with question–answer pairs and knowledge components, predicting the likelihood of correctly answering an exercise based on the estimated knowledge state. Similarly, preference tracing continuously estimates a user's preference state as they engage with content over time, predicting whether a user will enjoy a specific movie based on the estimated preference state. Our empirical evaluations demonstrate that Bayesian knowledge tracing (BKT)-based preference tracing not only delivers comparable predictive performance but also effectively captures users’ preference states at a component-wise level. Moreover, deep learning-based knowledge tracing (DLKT)-based preference tracing, which operates without predefined movie components, outperforms recent deep learning-based recommendation models, unveiling its potential to provide more accurate and nuanced recommendations.