Online Pattern Recognition in Multivariate Data Streams using Unsupervised Learning

Extracting patterns from data streams incrementally using bounded memory and bounded time is a difficult task. Traditional metrics for similarity search such as Euclidean distance solve the problem of difference in amplitudes between static time series prior to comparison by normalizing them. However, such a technique cannot be applied to data streams since the entire data is not available at any given time. In this paper, we propose an algorithm that finds patterns in data streams in an incremental manner by constantly observing only a fixed sized window through which the data stream flows. Our algorithm employs an unsupervised approach by which it randomly samples the data streams and eventually identifies good patterns by comparing samples with one another. It uses local information to find patterns globally over the data stream. The metric used for comparison is Euclidean distance between the first derivatives of the data points of the candidate patterns. Using such a metric eliminates the problem of different scales in the data streams. The output of the algorithm is a representation of the pattern in the form of normal distributions that characterize the ranges of values that might be observered. We show that our algorithm can be used to extract interesting patterns from both univariate as well as multivariate data streams.We measure the performance of our algorithm using standard metrics like precision and recall defined according to our context.