Impression prediction of VTubers using onomatopoeia for VTuber recommendation

Abstract

Recently, VTubers (Virtual YouTubers) are becoming more and more popular. While there are more than 20,000 VTubers, there are few ways to search for new VTubers with preference. A VTuber recommendation system considering VTuber's features is needed. There are few recommendation systems using video contents. Some studies discuss VTuber culture, however, no study analyzes video data of VTubers. Therefore, this research tries to predict the impression of VTubers' videos. The impression dataset of Vtubers is built by collecting 798 self-introduction videos of Vtubers and annotating them with onomatopoeia and Japanized Ten Item Personality Inventory (TIPI-J). VTuber impression vectors calculated from onomatopoeia and TIPI-J. The accuracy of the proposed impression model was not good, however, the results suggest that multi-modal processing is necessary to predict VTubers' impressions.