Clustering Time Series Online in a Transformed Space

Similarity-based retrieval has attracted an increasing amount of attention in recent years. Although there are many different approaches, most are based on a common premise of dimensionality reduction and spatial access methods. Relative change of the time series data provides more meaning and insight view of problem domain.. This paper presents our efforts on considering the relative changes of time series during the time series matching process. A similarity distance measure that based on transformed difference space of a series of critical points is proposed. Based on experiments with financial time series data, it can be concluded that our distance measure works as good as the Euclidean distance measure based normalized data without any shifting and scaling and PAA approach. The distance measure proposed is a general distance metric and is suitable to deal with online similarity matching because it does not maintain stream statistics over data streams.