Abstract
The advent of predictive methodologies has catalyzed the emergence ofdata-driven decision support across various domains. However, developing modelscapable of effectively handling input time series data presents an enduringchallenge. This study presents novel preference learning approaches to multiplecriteria sorting problems in the presence of temporal criteria. We firstformulate a convex quadratic programming model characterized by fixed timediscount factors, operating within a regularization framework. To enhancescalability and accommodate learnable time discount factors, we introduce anovel monotonic Recurrent Neural Network (mRNN). It is designed to capture theevolving dynamics of preferences over time while upholding critical propertiesinherent to MCS problems, including criteria monotonicity, preferenceindependence, and the natural ordering of classes. The proposed mRNN candescribe the preference dynamics by depicting marginal value functions andpersonalized time discount factors along with time, effectively amalgamatingthe interpretability of traditional MCS methods with the predictive potentialoffered by deep preference learning models. Comprehensive assessments of theproposed models are conducted, encompassing synthetic data scenarios and areal-case study centered on classifying valuable users within a mobile gamingapp based on their historical in-app behavioral sequences. Empirical findingsunderscore the notable performance improvements achieved by the proposed modelswhen compared to a spectrum of baseline methods, spanning machine learning,deep learning, and conventional multiple criteria sorting approaches.