Mining coherence in time series data

This paper presents work on modeling coherence between time series data. Work is based on the elaboration of formula that computes the distance between time series. Based on the computed distances, the method exploits the closest neighbor algorithm and leads to the construction of a phylogeny-clustering tree. Using car sales data (available on the www) we demonstrate work done and present preliminary results. We discuss implications of the work performed in correlating time series data with documents including such data. Introduction Time series occur in many aspects of economic and social activity. Time series organization of data implies time stamping of individual observations. Observations may represent economic or social activity or even life critical measurements such as those acquired when a patient is under monitoring in a critical care unit. Time series data modeling has been an active area of research in statistics. A variety of models exist, which manifest interest and provide the interested analyst with analysis tools (Box and Jenkins 1976). Expanding interest in data mining and knowledge discovery has contributed to an increase of research awareness in learning using time series data (Morik 2000). This paper reports preliminary research results and modeling activity in integrating time series data with documents. Motivation originates from the fact that analysts do use time series data to prepare reports. Reports and time series data reside in distributed information archives. There is an emerging need to mine for knowledge over such archives, which contain both reports and data. Specifically, the paper presents a methodology, which investigates similarity between time series data and reports results using exemplar time series drawn from car sales (see legend in Figure 1, in the text). Data-mining methodology is realized by the introduction of a novel algorithmic process and related formulas, for discovering similar and indicative patterns in time-series collections. Final outcome is the clustering of time-series into similar-groups, visualized by the appropriate customization of a phylogeny-based clustering algorithm and tool.