INQUIRY-BOUNDED MINING OF IMPRECISE DATA FOR ADAPTIVE INFORMATION MONITORING

Abstract

Information monitoring (IM) is an essential activity for management and control. Generally speaking, an IM system aims to continuously check for information updates. Once an update is detected, it triggers suitable actions such as exception validation, exception handling, and user notification. Since information sources (e.g., individual persons and servers) often cannot notify the IM system of updates, the IM system needs to actively inquire about the current status of the items being monitored. However, data inquiry may be costly, and uncontrolled inquiry may even exhaust the information sources. IM should thus be conducted under an inquiry-bounded constraint (i.e., limiting the number of inquiries). Under such a constraint, an IM system should inquire the right targets at the right time by adapting to the update behavior (i.e., rough update frequency at each time period) of each information item. This requirement brings a significant challenge to the design of the IM system: mining the imprecise data that is sampled in an inquiry-bounded manner. The data is imprecise in the sense that information sources often cannot indicate when, how many, and how frequently updates really happened. We tackle the challenge by developing a scalable multiagent model in which each agent performs autonomous mining and cooperative monitoring. It significantly outperforms state-of-the-art IM techniques in capturing a higher percentage of updates in a timelier manner by conducting fewer inquiries.