Development of Knowledge Discovery System from Annotated Time Sereis Data

Abstract

This paper proposes a knowledge discovery system from annotated time series data, they are expressed as sequences of numerical values. They generally have a lot of important information in background, but it is not included in data. Analysis methods without background information have limitations. Several studies propose meta data approaches, which is often expressed as a short text, the techniques are insufficient for our purpose. We therefore develop a method that uses annotations which are compact expressions of back ground information. Subsequences obtained by using domain knowledge are organized into groups based on a distance measure. Among the groups some of them are identified as important features. In order to measure importance, we develop a method that uses global and local frequencies of subsequences. This idea is similar to the TF*IDF method, which is used in text mining. A subsequence that represents an important group is regarded as a feature pattern. In addition, we extract association rules over feature patterns and annotations. We introduce a new concept, called max allowance length, to focus only on influential annotations to a pattern. We demonstrate an effect of the proposed method by using financial data.