Learning and Detecting Concept Drift

The volume of data that humans create has increased explosively as information science and technology have evolved. Therefore, the demand for learning machines that can extract input-output mappings and knowledge rules from massive data sets has become more urgent, and machine learning is now a core technology in the advanced information society. It has been applied to fields such as pattern recognition, search engines, medical support, robot engineering, image processing, and data mining and has achieved significant accomplishments in each field. Recently, several methods that could not be implemented with older computers have been developed with state-ofthe-art computers that have enormous memory capacity and high performance CPUs. Machine learning is expected to continue to develop in the future. Machine learning can be roughly classified into two types of learning based on how training examples are presented: batch learning and online learning. Batch learning systems are first given a large number of examples that they then learn all at a once. In contrast, online learning systems are given examples sequentially that they learn one by one. Many excellent batch learning systems have been proposed, but there are serious problems with online learning, especially in environments where the statistical properties of the target variable change over time. This change, known as concept drift, can happen either gradually or suddenly and significantly. The effectiveness of strategies for building a good learning system depends on the types of changes, so it is difficult to create an ideal learning system. A prime example of concept drift is the spam filtering problem. An effective spam filter must be able to handle various