Knowledge extraction from a time-series using segmentation, fuzzy matching and predictor graphs

In this paper, a novel multi-stage approach to knowledge extraction from a time-series is proposed. A given time-series is modeled as a sequence of well-known primitive patterns with the purpose of identifying first-order probabilistic transition rules for prediction. The first stage of the proposed model segments a time-series into structurally distinct temporal blocks of non-uniform length such that each block possesses a relatively low variation of dynamic slope. In the second stage, the temporal segments thus obtained are normalized and matched with one of four well-known primitive patterns using a fuzzy matching algorithm. Finally, the sequence of matched segments is used to represent the time-series as a set of four directed graphs corresponding to the four primitive patterns. Each vertex in the graphs represents a horizontal partition of the time-series and each directed edge indicates the transitions between two such partitions caused by the occurrence of one or more temporal segments. In the test phase, the graphs are employed to predict possible future values of the time-series. Experiments carried out on the TAIEX close-price time-series indicate a high prediction accuracy, thereby validating the use of the model for real-life forecasting applications.