MUKG: Unifying Multi-Task and Knowledge Graph Method for Recommender System

Abstract

Recommendation systems (RS) are automated mechanisms designed exclusively to prescribe desired products scientifically based on the preferences of the consumer. The users' interest would be either implied or precisely defined by the processes. Implicit meta-data collections merely provide all the possible detection of usage profiles. Recommendations may produce predominantly based on consumer anticipations, item features, user-interactions among other contextual factors. In popular RS techniques, collaborative filtering stoically endures from sparsity as well as cold start problems. To adequately address the key issues, many academic studies subtly manipulate feature learning to positively enhance the exceptional performance of RS. This paper ponders the Knowledge Graph to serve as the source of heterogeneous information. We introduce the MUKG a unifying framework based on multi-task feature learning and Knowledge graph to improve tremendously the recommendation. The MUKG represents a deep succeeding framework that appropriates the knowledge graph representation task to support the recommendation task. We develop a Mapping Layer Network and two inter-link tasks that learn precisely higher-order interactions between specific items and entities. We convincingly demonstrate MUKG represents a comprehensive framework incorporating many representative methods of RS along with multi-task learning. Through extensive experiments on four pragmatic datasets, we profusely illustrate that the potency of the MUKG over different newfangled touchstone. Precisely, our developed model properly obtains average Accuracy accumulations of 11.6%, 11.5%, 12.7%, and 8.7% in movies, books, music and news recommendation, individually.