The Impact of Noisy Information in Knowledge Graphs on Recommendation Performanc

Abstract

Knowledge graphs (KGs) have been widely used in recommender systems (RSs) as item auxiliary descriptions for high-quality recommendation. In current KG-based RSs, KGs are usually built based on entity linking and name matching operations. The limited manual supervision during the construction process will produce the untrustworthy information in KGs. In addition, entities in KGs suffer from long-tail distribution problem and contain connections that are irrelevant to the recommendation target. Such untrustworthy information and irrelevant connections is noise in KGs and becomes an obstacle to high-quality recommendations. In order to clearly show the impact of noisy information in KGs on recommendation tasks, we propose a general way to effectively remove these noises from knowledge graphs. Furthermore, we combine our method with current KG-based methods, and the improvement in recommendation performance shows the harm of noise information in KGs to recommendation tasks. It also clearly demonstrates the necessity of current KG-based RSs to detect and remove noise information in KGs.