GRAPH BASED FRAMEWORK FOR TIME SERIES PREDICTION

Abstract

A time-series is a sequence of real numbers where each number represents a measured value with respect to time. Data mining when performed on time series data is called time series data mining. Time series data mining comprises of tasks like clustering, dimensionality reduction, association rule mining, classification and prediction. A major task of data mining with regard to time series is predicting the future values. The time series prediction can be useful in many ways such as planning, measuring the performance of a predicted value based on the past data against actual observed value. Existing methods of time series prediction like fitting a function, use of least square method, exponential method and ARIMA method are statistical approaches which require experience and prior knowledge by the user. These methods also do not consider about evolving patterns in time series over periods of time. Thus there exists a major research challenge of dealing with evolving patterns in time series. We propose a new approach for time series prediction that is based on graph framework and takes care of evolving patterns in time series data. This framework overcomes the problem by extracting concealed patterns in time series using concept of graph and graph mining techniques. The motivation behind using graphs on time series data is that graphs can model each observation as vertex and represent the affect of variation in observations with respect to time in form of edges. We, thus propose to map time series data to graphs and use graph edit distance measure for computing similarity in time series data being represented as graphs. This similarity is being used for purpose of classification and prediction. In the present work, the