Adaptive Information Filtering Based On Incremental Learning And Threshold Optimization

With the amount of online information growing rapidly, people were surrounded by the information overload problem. Information filtering focuses on this issue and retrieve information relevant to the users' specific requirements. In order to meet web-based on-line time-critical information filtering needs, adaptive information filtering was paid more attention by researchers, which requires little information about users' interests to construct the profile. Along with the filtering process, adaptive information filtering can learn actively form user's periodic feedback and adapt itself to produce the high performance. The main work in this dissertation is to study the new algorithm of profile learning and new approval to optimize the dissemination threshold.The paper proposes an improved adaptive information filtering model, which combined the statistical model (vector space model) with probability approach (Bayesian inference method). Documents of incoming stream and user topic description were represented and matched in the vector space model. Initial profile was constructed by the vector median method. New Machine learning algorithm was designed to get the accurate profile via the pseudo relevant feedback. Gaussian and exponential distribution model was set to get the document probability of relevant over training data which afforded from the user periodic feedback. Bayesian inference based on linear utility function to exploration the optimization threshold.At the training step of model, incremental learning algorithm using pseudo relevant feedback was presented in this dissertation. Feature selection method based on document frequencies was introduced to pick the informative term alone with the profile learning. On the test step, new method was introduced to exploration the dissemination threshold. An incremental Rocchio algorithm was introduced to adapt profile vector based on the user periodic feedback information.New model received a good performance on the Chinese corpus. Experiment results indicate that incremental learning and threshold optimization are effective in the adaptive information filtering.