Representação vetorial consciente de tempo para recomendadores incrementais

Abstract

With the growth of data volume, recommendation systems have evolved from static batch approaches to incremental models, allowing intelligent agents to treat recommendation as a dynamic decision process and learn continuously from user feedback. However, by prioritizing stages such as policy optimization and reward modeling, the current literature neglects important aspects of these systems, such as context representation and temporal consumption dynamics, factors intrinsically related to the continuous recommendation scenario. Based on this gap, this work introduces TAI2Vec, an item embedding generation model that integrates temporal variables to create latent representations adaptable to different consumption rhythms. Through an extensive experimental evaluation, covering from similarity recommendations to incremental learning algorithms based on bandits, it is demonstrated that TAI2Vec surpasses traditional methods in recommendation tasks, especially in scenarios with data scarcity, evidencing the capacity that temporal information has in enriching contextual learning in different recommendation approaches. The source code for the TAI2Vec methods is publicly available at: https://github.com/UFSCar-LaSID/tai2vec.