Association Rules Mining of Existentially Uncertain Data

Speaker: CHUI Chun Kit

Abstract

We study the problem of mining association rules from existentially uncertain data under a probabilistic framework. In many application areas such as pattern recognition systems, information retrieval and click streams analysis, object features are often expressed in terms of probabilities, or the likelihood of their presence . Extracting association rules among these features requires heavy computational and I/O efforts. Since each feature of an object is associated with an existential probability, conventional association rule mining algorithms become either inapplicable or unacceptably inefficient. We present a simple theoretical framework for mining association rules from existentially uncertain data and discuss the applicability of traditional association rules mining algorithms . A Preliminary study is conducted to identify the computational bottleneck of the mining problem. We propose a number of efficient methods to reduce both computational and I/O costs in the mining process . Experimental results show that our methods can achieve significant computational and I/O saving.