Magnitude Inference Method Using Inductive Inference

A Case Study Based on the Seismicity Patterns of Intraplate Earthquakes in Japan

Abstract

   With the development of research on the machine learning technique, several methods of inductive inference have been proposed. Among them are methods that use a decision tree which is inductively constructed for classifying given examples efficiently. In this paper this method is applied to inferring the magnitude class of intraplate earthquakes in Japan by using the seismicity patterns cataloged by Ichikawa (1986). Three methods are introduced here for making a decision tree. The first method, which was originally proposed by Quinlan (1983), uses the entropy as the index to select the node parameters of a decision tree. The second is a new method based on the AIC (Akaike Information Criterion) instead of the entropy. The third one, called here the NAIC method, uses only the node parameters whose AIC value is negative. To compare these three methods the alarm rate and truth rate of inference are calculated for two kinds of data, one containing noise and the other containing no noise. The NAIC method was found to give the highest value for both the alarm rate and truth rate whether the data contain noise or not. These methods do not work well for earthquakes of magnitude 7 class. This is because the frequency of earthquake occurrence of magnitude 7 class is very low, and information on it is not sufficient to get a reliable decision tree.