Web multimedia data mining
We are trying to develop new applications which fully utilize rich information in multimedia, using real big data on Web.
Content-based recommender system applying a generic image recognition methodology
In the field of recommender systems, the cold-start problem, which is a difficulty in recommendations regarding users and items with little or no historical data, is one of the most fundamental problems. We are developing a content-based recommendation approach which applied a generic image recognition methodology to overcome the cold- start problem. In our method, image features extracted from items by Convolutional Neural Networks (CNN) are used as explanatory variables, and desired meta-data such as user profiles and click through rates (CTR) are used as objective variables to construct classifiers. These classifiers enable us to make recommendations reflecting the content of items. Moreover, our proposed method is completely content-based using image similarity; therefore, we expect it to enable recommendations even for users and items with little historical data. As applications, we are developing an estimating system of the target viewership of video clips on video sharing sites and a recommender system of display advertisements.
- Kohei Yamamoto, Riku Togashi, Hideki Nakayama, "Content-based viewer estimation using image features for recommendation of video clips", ACM RecSys 2014 Workshop on RecSysTV, 2014. pdf