Prediction in dynamic system - a divide and conquer approach

Abstract

The aim of this work is to find a general framework for making decision from large and complex dynamic data, where all the influencing attributes are not known, and some of the available, apparently relevant, attributes could really be irrelevant. Some examples are: various diagnostic data from patient monitoring system, data from financial market, or the sales data of a big chain store for analyzing the buying pattern of customers. In these systems, many factors interact in a complex manner so that a complete analysis is often impossible, and conventional statistical methods for prediction also fail. In this work, we propose a framework for prediction in such problems using soft computing tools. We find partitions in the multivariate space so that data of same targeted forecast or decision are grouped there to the extent possible. This search is performed at different subspace level, i.e., a feature-subset selection. This part is accomplished by using rough set theory, after discretizing the original data. We conjecture that, if sufficient number of data with same decision falls in one of those subspace partition, any data in that partition would be predictable. A data can be member of more than one such subspace partition. In the next step, we train individual neural networks for each such subspace partition to learn the input-decision mapping using the original continuous valued data belonging to that subspace partition. For a new data, the ensemble of the trained neural network expert systems takes the decision. When applied to the prediction of stock-value, our model gave better results compared to other methods.