Human-Centered Explainable Recommender System Using Cognitive Preference Drift Modeling
TL;DR
CPD-XRS (Cognitive Preference Drift Explainable Recommender System) is introduced that models psychologically sound processes of preference drift including gradual drift, sudden drift, cyclic drift, hedonic adaptation, and mere-exposure effect.
Abstract
The traditional approach to recommendation systems makes a static assumption about the user’s preferences that does not have scientific support from psychology. This paper introduces CPD-XRS (Cognitive Preference Drift Explainable Recommender System) that models psychologically sound processes of preference drift including gradual drift, sudden drift, cyclic drift, hedonic adaptation, and mere-exposure effect. We use a temporal variational autoencoder to model drift detection, psychology-inspired transition model for modeling drift according to Mere-Exposure and Hedonic Adaptation equations, and knowledge graph-based explainability technique that allows to make the method understandable and human-centred. Our experiments on Amazon Books (18 years, 22.5M interactions) and MovieLens-25M (25 years, 25M ratings) datasets demonstrate CPD-XRS performance in terms of NDCG@10 score of 0.4012 and 72.1% accuracy of drift detection which outperforms state-of-the-art baselines TimeSVD++, CDR [5] and PPD [2]. The user study with 120 participants demonstrates the ability of our method that uses timeline visualization on explanations to produce trust (4.31/5.00), satisfaction (4.18/5.00) and perceived fairness (4.09/5.00) which is significantly higher compared to baselines without explanation or textual explanations. The presented work merges psychological theories with recommendation systems demonstrating that modeling of cognitive drift is not only theoretically sound but empirically advantageous.