Data Stream Classification by Adaptive Semi-supervised Fuzzy Clustering

The analysis and classification of data streams has attracted much attention recently due to the increasing amount of applications that produce streaming data. Most of the existing work relevant to data stream classification assume that all data are completely labeled. However in many applications, labeled data are difficult or expensive to obtain, meanwhile unlabeled data are relatively easy to collect. Semi-supervised learning algorithms can solve this problem by using unlabeled samples together with a few labeled ones to build classification models. In [1] we introduced a data stream classification method based on an incremental semi-supervised fuzzy clustering algorithm. The method assumes that data belonging to different classes are continuously available during time and processed as chunks. The clusters are formed from a chunk via the SSFCM (Semi-Supervised FCM) clustering and when the next chunk becomes available the clustering is run again starting from cluster prototypes inherited from the previous chunk. The algorithm creates a fixed number of clusters that is set equal to the number of classes. In real-world contexts the underlying distribution of data may change over the time, hence a fixed number of clusters may not capture adequately the evolving structure of streaming data. To overcome this limitation in this work we extend the method proposed in [1] by introducing the capability to adapt dynamically the number of clusters. When the cluster quality deteriorates from one data chunk to another, the number of clusters is increased (by splitting some clusters) or also decreased (by merging some clusters). The cluster quality is evaluated in terms of the reconstruction error [2] that measures the difference between the original data and their "reconstructed" counterpart derived using the clustering outcome (prototypes and membership degrees). Preliminary experimental results on the benchmark data set KDD-CUP99 show that the proposed adaptive version of the data stream classification method outperforms the previous static version and is more robust in presence of outliers.