The main thrust of this paper is application of a novel data mining approach on the log of user’s feedback to improve web multimedia information retrieval performance. A user space model was constructed based on data mining, and then integrated into the original information space model to improve the accuracy of the new information space model. It can remove clutter and irrelevant text information and help to eliminate mismatch between the page author’s expression and the user’s understanding and expectation. User space model was also utilized to discover the relationship between high level and low level features for assigning weight. The authors proposed improved Bayesian algorithm for data mining. Experiment proved that the authors’ proposed algorithm was efficient. The main thrust of this paper is application of a novel data mining approach on the log of user’s feedback to improve web multimedia information retrieval performance. A user space model was constructed based on data mining, and then integrate into the original information space model to improve the accuracy of the new information space model. It can remove clutter and irrelevant text information and help to eliminate mismatch between the page author’s expression and the user’s understanding and expectation. User space model was also utilized to discover the relationship between high level and low level features for assigning weight. The authors proposed improved Bayesian algorithm for data mining. Experiment proved that the authors’ proposed algorithm was efficient.