Introduction

One of the most difficult problems with classification is that the concepts and their relations are highly dependent on some hidden context. As consequences, any information about number of concept, description of concepts, relation between the concepts should ideally be totally well-known. Most often, a priori or arbitrary knowledge is injected in the learning scheme to simplify the task and make it more accurate. This is the reason why many supervised classification tasks were done under the assumption that the contexts are fixed and known before the learning. Recently, a great attention has been paid on classification in data streams where the concepts can change over the time. The classification of data delivered in stream mainly use very specific techniques and adaptations. Indeed, for many real processes, the categorization of the data is not so simple, even in a supervised mode. Each time the user encounters a new data, he can decide to change the previously established classification, for example by merging classes, creating new ones and so on. More generally, the concepts can evolve over the time depending on encountered data : it can lead to concepts fusion or splitting, to concepts creation, extension or deletion. The knowledge and expertise are not precisely accessible at the beginning of the process mainly because final needs and usages can vary.

 This multi One-Class incremental SVM (so called mOC-iSVM) system has been proposed as a solution to learning problems in a context of data streams. In the field of incremental learning, different challenges are addressed to adapt the system to stationary (fixed concepts) or non-stationary environments (concept drift), also challenged with adding, removing, merging, splitting classes. The mOC-iSVM has shown its ability to improve the classification accuracy over the time, with incoming new labeled data, without performing any complete retraining.