Dynamic learning of cases from data streams

This paper presents a dynamic adaptive framework for building a case library being able to cope with a data stream in the field of Case-Based Reasoning. The framework provides a three-layer architecture formed by a set of case libraries dynamically built. This Dynamic and Adaptive Case Library (DACL), can process in an incremental way a data stream, and can be used as a classification model or a regression model, depending on the predicted variable. In this paper, the work is focused on classification tasks. Each case library has a first layer formed by the dynamic clusters of cases, a second one formed by the meta-cases or prototypes of the cluster, and a third one formed by an incremental indexing structure. In our approach, some variant of k-d tres have been used, in addition to an exploration technique to get a more efficient retrieval time. This three-layer famework can be constructed in an incremental way. Several meta-case learning approaches are proposed, as well as some case learning strategies. The framework has been tested with several datasets. The experimental results show a very good performance in comparison with a batch learning scheme over the same data.