Feedback-Based Performance Tuning for Self-Organizing Multimedia Retrieval Systems

We concentrate on content-based retrieval in unstructured P2P networks consisting of thousands of peers that unpredictably join and leave the network. Such environments with permanent churning of peers require self-organizing mechanisms that should deal with sudden peer failures, arrivals of new peers, and continual changes of data or network topology. In this paper, we propose a self-organizing search system that operates in an unstructured P2P network and allows users to search for multimedia data by their content. The peers are interconnected by relationships established according to answers returned to queries. In order to select appropriate relationships for query forwarding, we define and evaluate a new adaptive routing algorithm. The routing algorithm is influenced by automatically evaluated feedback, so the system does not need any user intervention. The experiments evaluated on a real-life image data set confirm that the proposed technique achieves high recall values, while contacts a moderate number of peers. In addition, we demonstrate resilience of the system to sudden peer failures by studying system performance and a quality of returned answers after a large number of peers is disconnected.