Adaptive Filtering for Information Gathering WWW Agents

Motivated by the shortcomings of large, centralized WWW information services, we focus on improving the performance of personalized WWW information gathering agents. We point out the strong connection between text ltering and WWW information agents, and implement a set of ltering algorithms intended for use with such agents. We propose an objective evaluation method for WWW information agents based on IR benchmarks and the desired criteria for WWW agents, and present a comparative study of the implemented ltering algorithms. From our experiments, we conclude that positive examples of relevant documents are suucient for training a ltering algorithm , that users' topics of interest can be learned and are not needed as part of the training process, and that performance degrades gracefully when the amount of training provided to the lter is limited.