Adaptive Lightweight Text Filtering

We present a lightweight text ltering algorithm intended for use with personal Web information agents. Fast response and low re- source usage were the key design criteria, in order to allow the algorithm to run on the client side. The algorithm learns adaptive queries and dis- semination thresholds for each topic of interest in its user prole. We describe a factorial experiment used to test the robustness of the algo- rithm under dieren t learning parameters and more importantly, under limited training feedback. The experiment borrows from standard prac- tice in TREC by using TREC-5 data to simulate a user reading and categorizing documents. Results indicate that the algorithm is capable of achieving good ltering performance, even with little user feedback. Text ltering makes binary decisions about whether to disseminate documents that arrive from a dynamic incoming stream. Adaptive ltering systems (3) start out with little or no information about the user's needs, and the decision of whether to disseminate a document must be made when the document becomes available. The system is given feedback on each disseminated document to update its user prole and improve its ltering performance. In this paper, we present a lightweight ltering system designed for use in personal Web information agents. We assess how well the algorithm works with little feedback, a requirement for its application as a personal information gath- ering agent. Also, we assess the eect of algorithm parameters on robustness. We make two key contributions in our research. First, our algorithm adapts standard ltering techniques to the needs of personalized web applications: lightweight, privacy protecting and responsive to user provided examples. The algorithm learns a prole of user information interests. Second, we adapt a rig- orous evaluation method to web systems by using text ltering benchmarks to simulate user behavior. Traditionally, Web systems have been evaluated with user studies, with the disadvantages of slow data collection, little experimental control and decreased objectivity of conclusions. Relying on simulated user feed- back allows us to test many alternative design decisions before subjecting the system to a user study, which means we are less likely to waste subjects' time and are more likely to produce a well tuned system.