Automated Multimedia Systems Training Using Association Rule Mining

User feedback is widely deployed in recent multimedia research to refine retrieval performance. However, most of the existing online learning algorithms handle interactions of a single user, which may pose restrained performance due to the limited size of positive feedback. An alternative solution is to learn general user perceptions via collecting feedback from different users. The training process is initiated only when the number of feedbacks reaches a certain threshold. This could improve the performance but it becomes a manual process to decide the threshold and initiate the training process. To address this challenge, ire propose an advanced training method by adopting the association rule mining technique, which can effectively evaluate accumulated feedback and automatically invoke the training process. Training is performed per video rather than for all videos in the database, making the process more efficient and robust. In addition, it can further improve semantic modeling in the video database and continuously improve retrieval performance in the long run. As an example, the proposed method is applied to a mobile-based soccer video retrieval system and the experimental results are analyzed.