A user study with aspect-based sentiment analysis for similarity of items in content-based recommendations

Most studies on recommender systems focus on collaborative algorithm approaches over content-based recommendation due to their better accuracy results. However, the advantage of the latter is that it is more effective and more transparent with user applications. This article proposes WordRecommender, an explainable content-based algorithm that calculates similarity by semantic proximity. Its preprocessing step involves analyses of movie reviews to obtain aspects, defined as relevant words of high sentimental value. Recommendations are generated by a neighbourhood algorithm that calculates the similarity of films based on the semantic proximity of the aspects ordered by their emotional score. It can also consider a semantic comparison of metadata using the most related aspects from the recommended movie and one enjoyed by the user for the production of textual explanations. The accuracy of the algorithm was competitive with those of other baseline neighbourhood methods, and the semantic data of items can be the source of both representative information and reasoning in recommender systems.