Incremental learning and concept drift: Editor's introduction

A complex problem in data analysis is the time-varying nature of many realistic domains. In many real-world learning problems, training data become available in batches over time, or even flow steadily, as in user-modeling tasks, dynamic control systems, web-mining, and times series analysis. In these applications, learning algorithms should be able to adjust the decision model dynamically whenever new data become available. This is the scenario that motivates this special issue of Intelligent Data Analysis to machine learning systems capable of dealing with concept drift. To narrow the domain of interest, we focus on those learning scenarios in which the system must induce the concept from timestamped training data. A brute-force algorithm relearns the concept from scratch each time a new example becomes available. This poses several problems. Learning from the scratch wastes computational resources. Moreover, in non-stationary environments, the system should take into account the fact that only the most recent examples are relevant to the actual target concept. A less expensive approach would employ an incremental learning technique that adapts the previously induced concept model by incorporating the experience obtained from newly available examples. An incremental learning system can be used with some success for domains in which the underlying instance distribution evolves, especially if there is an abundance of examples that are representative of the most recent version of the target concept. Yet, in domains where the change is substantial and there is a paucity of recent examples, the system needs to be able to discount or even forget older examples, and adjust what has been induced from them. The task is more difficult than it appears. When learning in time-varying domains, the system needs to modify the internal concept representation not only as more examples become available, but also in response to suspected changes in the definition of the target concept. It is of paramount importance that the system be able to distinguish between the situation in which new examples only help to fine-tune the existing concept model, and the situation in which the new examples are indicative of a shift in the target concept. To complicate matters even further, the system should not be misled by noise. Over the past decade, many researchers have become interested in this task, and the results of their work have appeared in diverse journals and conferences. By organizing this special issue, we wanted to concentrate several alternative approaches in the same volume in order to give the interested reader a better idea about the state-of-the-art of the relevant algorithms, applications, and evaluation methods. We believe that the five articles that appear here satisfy this goal.