An incremental learning algorithm with extended data expression

This paper introduces an ensemble system that handles data containing probabilistic concepts. This system is to construct a data structure automatically by applying a data extension method which has advantages to be applied to various application fields and to produce a good performance result.

The extended data representation method is designed to introduce the concept of probability into general data. In this method, weights can be assigned to each event so that the value of each event can be different. Also, attribute value is divided by cardinality so that each attribute value can be represented stochastically, and this value is expressed as a probability value between 0 and 1. This paper introduces UChoo algorithm which modified C4.5 to this expression method and introduces AdaBoost algorithm which uses UChoo as weak learner. The most representative of the Boosting methods, AdaBoost, assigns a weight to each event in the training data. This weight increases the value for misclassified events and controls the probability of inclusion in the next learning data set. It is known that this algorithm converges quickly to the target and combines weak learner without the need for special parameters to produce good results. Therefore, extended data representation composed of weight and probability is considered to be suitable for the AdaBoost series algorithm.

The experimental results confirm that this algorithm can learn normally without incurring too much loss by receiving incomplete data. When collecting data from different environments and sensors in ubiquitous environment, some data is damaged or incomplete data is generated by using different sensors. At this time, these advantages may be useful for such incomplete data processing fields.