Tiered Clustering for Time Series Data

Abstract

Clustering is an essential unsupervised learning method. While the clustering of discrete data is a reasonably solved problem, sequential data clustering, namely time series data, is still an ongoing problem. Sequential data such as time series is widely used due to its abundance of detailed information. Often, normalization is applied to amplify the similarity of time series data. However, by applying normalization, measurement values, which is an important aspect of similarity, are removed, impairing the veracity of comparison. In this paper, we introduce a tiered clustering method by adding the value characteristic to the clustering of normalized time series. As such, two clustering methods are implemented. First, the Distance Density Clustering algorithm is applied to normalized time series data. After obtaining the first-tier results, we apply a traditional hierarchical clustering of a summarized time series value to further partition clusters.