Extraction de motifs séquentiels dans les flux de données. (Sequential patterns mining from data streams)

In recent years, many applications dealing with data generated continuously and at high speeds have emerged. These data are now quali ed as data streams. Dealing with potentially in nite quantities of data imposes constraints that raise many processing problems. As an example of such constraints we have the inability to block the data stream as well as the need to produce results in real time. Nevertheless, many application areas (such as bank transactions, Web usage, network monitoring, etc.) have attracted a lot of interest in both industry and academia. These potentially in nite quantities of data prohibit any hope of complete storage ; we need, however, to be able to examine the history of the data streams. This led to the compromise of "summaries" of the data stream and "approximate" results. Today, a huge number of di erent types of data stream summaries have been proposed. However, continuous developments in technology and in corresponding applications demand a similar progress of summary and analysis methods. Moreover, sequential pattern extraction is still little studied : when this thesis began, there were no methods for extracting sequential patterns from data streams. Motivated by this context, we are interested in a method that summarizes the data stream in an e cient and reliable way and that has as main purpose the extraction of sequential patterns. In this thesis, we propose the CLUSO (Clustering, Summarizing and Outlier detection) approach. CLUSO allows us to obtain clusters from a stream of sequences of itemsets, to compute and maintain histories of these clusters and to detect outliers. The contributions detailed in this report concern : Clustering sequences of itemsets in data streams. To the best of our knowledge, it is the rst work in this domain. Summarizing data streams by way of sequential pattern extraction. Summaries given by CLUSO consist of aligned sequential patterns representing clusters associated to their history in the stream. The set of such patterns is a reliable summary of the stream at time t. Managing the history of these patterns is a crucial point in stream analysis. With CLUSO we introduce a new way of managing time granularity in order to optimize this history. Outlier detection. This detection, when related to data streams, must be fast and reliable. More precisely, stream constraints forbid requiring parameters or adjustments from the end-user (ignored outliers or their late detection can be detrimental). Outlier detection in CLUSO is automated and self-adjusting. We also present a case study on real data, written in collaboration with Orange Labs.