Improving human-computer interaction in personalized TV recommender

Abstract

In today's world of numerous sources of multimedia content, recommender systems help users find relevant content items. In our research the reasoning behind the recommendations generated by such systems was explored to check whether presenting users with explanations of recommended content increases their trust in the system. A content-based recommender for TV content has been developed which focuses on items attribute values. The system predicts users' ratings by classifying the vector of similarities between the user model and the items attributes. Users' trust is increased by identifying attribute values that are the most relevant for them. Users' feedback to the identified attribute values was used to improve the performance of the recommender algorithm. Tests in our experimental platform showed that the developed algorithms produce good results. The accuracy of the system was around 75% in the basic version and it further increased in the enhanced, while the identification of relevant attribute values achieved 86% precision.