Feedback-Driven Pattern Matching in Time Series Data

While motif discovery methods have come a long way over the years, they generally match occurrences based on the similar shape of the whole subsequence. As patterns in production network monitoring environments, which monitors and manages entire infrastructures of millions of device metrics over time, frequently exhibit more complex characteristics such as differences in temporal size or expected noise, these methods often remain insufficient for accurately tracking important behavioral patterns such as backup cycles or transcode sessions. This paper therefore proposes a feedback framework that allows a user to select additional motif ranges to be included or excluded from the model. The method uses a distance matrix of subsequences to extract common patterns from feedback samples and defines temporal rules on how these patterns are allowed to occur. The technique was tested on synthetic data as well as production network monitoring data and a publicly available human motion primitives dataset. The tests show that the recall score can be significantly improved with the proposed feedback system, increasing from 37% to 95% while also maintaining a perfect precision score. This is achieved by providing only one to three feedback samples as input. While the scope of this paper is limited to shape based features, the proposed technique can also be used for less exact patterns such as changepoints in noise.