Multivariate Anomaly Detection Using Frequent Itemset Mining

Widespread use of IoT devices that is predicted to generate 79 zettabytes of data annually by 2025 is only one example to emphasize the importance of time series data mining, especially anomaly detection both in academia and industry. Despite extensive research producing hundreds of algorithms, current frameworks inadequately abstract technical complexity for domain experts. In this dissertation, we introduce a novel framework that efficiently aggregates and summarizes anomaly scores from very high-dimensional datasets comprising millions of data points, enabling flexible query support and precise responses. Our approach employs a windowing technique to transform multidimensional anomaly scores into a transaction database, thereby leveraging established itemset mining algorithms. We use Matrix Profile, a well recognized methods to detect discords, although theoretically this approach can use any anomaly scores. Experimental results on real-world as well as large-scale synthetic datasets demonstrate that our method outperforms existing similar multidimensional approaches in both speed and query capability. To the best of our knowledge, this is the first work to integrate multidimensional anomaly detection with itemset mining. We illustrate its potential through several application scenarios and provide a modular design that invites further extension by the research community.