A New Framework for Data Streams Classification

Mining data streams has recently become an important and challenging task for a wide range of services, including credit card fraud detection, sensor networks and web applications. In these applications data do not typically take the form of persistent relations, but tend to arrive in multiple, continuous, rapid and timevarying data streams. Hence, conventional knowledge discovery tools cannot manage this overwhelming volume of streaming data. The nature of data streams requires the use of algorithms, which involve at most one pass over the data, and try to keep track of time-evolving features, also known as concept drifting. In the literature, several techniques for the efficient computation of compact space representations of massive data have been proposed to improve model reliability, including methods to cope with concept drifting. The challenges of these methods are that they are allowed to use small space and time to process a single item, while they must provide an accurate representation of some relevant characteristics of data streams. Based on the above considerations, our approach attempts to solve the problem of data streams classification, presenting a framework based on two main components. First, an on-line component is used to compute and manage some statistical aggregates information about data streams. The latter are viewed as divided into chunks, and an aggregate representation, called snapshot, is defined and built out of each stream chunk. All the statistics produced are stored in a summary structure called frame, using different levels of granularity. Second, a mining component is employed to extract a classifier from all the snapshots produced. In order to maximize the number of elements from which the model is extracted, this component employs snapshots with different levels of granularity, while only most recent data are used in the presence of concept drifting. We argue that the proposed approach avoids the continuous update of the same mining model, i.e. new models are continually extracted from scratch, while the system constantly updates summarized data structures. Furthermore, this approach employs a selective ensemble classifier combining the predictive power of the models extracted with different and increasingly larger data chunks representing different portions of data streams. Since both the performance and the accuracy of the whole system are based on single model evaluation, it is necessary to check the set of actually activatable models to classify new instances,