グラフニューラルネットワークを用いたコンテンツ推薦における知識表現の実証分析

抄録

In recent years, entertainment content, such as movies, music, and anime, has been gaining attention due tothe stay-at-home demand caused by the expansion of COVID-19. In the content domain, research in the field ofknowledge representation is primarily concerned with accurately describing metadata. Therefore, different knowledgerepresentations are required for applications in downstream tasks. In this study, we aim to clarify effectiveknowledge representation for predicting users’ latent preferences through a case study of an anime recommendationtask. We developed hypotheses from both quantitative and qualitative aspects on how to represent work knowledgeto improve recommendation performance, and verified them by changing the structure of knowledge representationaccording to the hypothesis. Initially, we constructed a Knowledge Graph (KG) by integrating domain-specific andgeneral-purpose data sources through the process of entity matching and imposing constraints on the properties.Subsequently, we constructed multiple KGs by varying the knowledge configuration. Specifically, we changed thecomposition of the data sources considered in the KG construction or excluded a triplet associated with an arbitraryproperty. After that, we fed the constructed KGs into the graph neural network recommender model and comparedthe recommendation performance. As a result, it was shown that the recommendation performance based on the KGcomposed of multiple data sources was the best, thus supporting the hypothesis from a quantitative aspect. Next,an ablation study on the properties revealed that knowledge characterizing the work itself contributed to the recommendationperformance, thus supporting the hypothesis from a qualitative aspect. Furthermore, we constructed atext-based KG by generating a new vocabulary from the “synopsis” text. It can describe the work’s storyline andworldview in more detail. We take it as an input to a Large Language Model (LLM) and extend the existing metadatabasedKG. The results showed that the KG considering both metadata and text had the best overall recommendationperformance, again confirming the hypothesis.