Detecting Novelties by Mining Association Rules

We propose a novelty detection method based on association rule learning as a candidate approach for validating online adaptive systems. Support and confidence intervals in association rules are used as a basis for examining a tested example to trace abnormality. Using a simple mutation algorithm for abnormality generation, the method has been evaluated on different data sets. Obtained experimental results are presented in the paper. Since it is based on association rule learning, our approach scales to very large data sets. A notable advantage of this method is that novelty detection thresholds can be obtained from empirical testing, rather than being predefined.