A new method of mapping relations from data based on artificial neural network

Abstract

Many methods have been proposed to derive patterns and correlations from data, such as artificial neural network (ANN) and support vector machine. They successfully learned the pattern from data but failed to illustrate them to human. In this paper, a new method which reveals the influences between factors and identifies key correlations among them from ANN is proposed. The method extracts the relations as relation maps, which is a perceptive illustration to interpret the actual logic beneath the neuron matrices. In this paper, we propose some definitions, express our method in mathematics, and develop a simplified practical algorithm of our method. Then we apply our algorithm to weather forecast problems. The algorithm successfully excavates the relations among weather factors, maps the relations to explicit graph, identifies the key relations, and thus used to reduces the inputs of predicting ANNs by 60 %. Generally, the pruned ANNs performed effectively. Of all the 14 predicting ANNs, when pruned, 10 have prediction errors no more than 16 % greater than the original ANNs, and 5 of the 10 have even lower prediction errors than before. Such result shows that our method successfully identifies key relations among factors. This is a justification for the reliability of our method on extracting relations encoded in ANNs. This promising method can be widely applied in the field of data mining and knowledge discovery.