Cross-Domain Recommendation for E-Commerce Services
In recent years, e-commerce services such as Amazon and Rakuten have been offering product recommendations in the form of “recommendations for you.” By recommending products to users, the burden on users of searching for desired products among a huge number of products can be reduced. In the current studies on recommendation systems, users and products are nodes, graphs with purchase histories and product evaluations as edges are constructed, and graph analysis methods are implemented to recommend products. However, in these studies, there is a problem that the accuracy of recommendations through graph analysis decreases for new users and products with little purchase history and evaluation. In the Laboratory of Media Dynamics, we are exploring a cross-domain recommendation method that can solve the cold-start problem by applying knowledge extracted from users and products with rich purchase histories and ratings to new users and products.
Highlight Video Generation from Sports Broadcast Video Using SNS
Highlight video generation is a video summarization technique that extracts the most exciting and conspicuous scenes from videos of sports and TV programs. By watching the generated highlight video, viewers can efficiently identify only the important scenes in the video. Since important scenes have visual or auditory features that differ from those of other scenes, previous research has used information obtained from the video to generate highlight images. However, information obtained from only videos is insufficient to generate highlight images from videos such as sports videos, where expert knowledge is required to determine important scenes. Text posted on social networking sites, on the other hand, contains information posted by viewers who have expert knowledge. Thus, in the Laboratory of Media Dynamics, we are conducting research on a method for generating highlight videos across different types of media by utilizing text information obtained from social networking services, along with information obtained from videos. Specifically, we are extracting information about the target sports video from texts posted by viewers on Twitter and analyzing it to extract scenes according to viewers’ preferences.
Music Recommendation Based on Reinforcement Learning for Music Distribution Services
With the prevalence of music distribution services such as Spotify and YouTube Music, an enormous amount of music is available on the Internet. To help users discover their favorite songs, various online services have introduced song recommendation systems based on users’ preferences. However, in current music recommendation systems, the process that leads to the recommendation result is unclear, and users cannot understand the relationship between the recommended music and their preferences. Therefore, in the Laboratory of Media Dynamics, we are exploring a music recommendation technology that can explain the process leading to the recommendation result. Specifically, we search for songs that match a user’s preferences exhaustively through reinforcement learning and present a search path to the user to achieve highly explainable song recommendations.