Abstract
Abstract
The streaming data from different resources as social media, telecommonication network or credit cards are accumulated and growing enourmously. Thus, it has become more important to produce valuable information from such big data environments. There are some characteristics of data streams such as continuous flow, high volume, rapid arrival and change of distribution. Due to these characteristics, there are some limitaions for processing data streams such as limited resouece and time and the data can be read only once. At this point data stream mining emerges with the streaming version of traditional data mining operations such as clustering and classification. In this study, we worked on data stream classification and short text stream clustering as a specific area of data stream clustering. We proposed some enhancements and new methods and compared the performance of these ones with the state of the art methods. For data stream classification, we named our enhancements as m-kNN (Mean Extended kNN) which can be considered as a derivative of traditional kNN (K Near- est Neighbours) and CSWB (Combined Sliding Window Based) classifier which is an ensembler and the combination of m-kNN and MC-NN (Micro Cluster Nearest Neighbour). We also present two new versions of CSWB, CSWB-e and CSWB-e2 such that our m-kNN classifier is combined with K* (K-Star) and C4.5, and with K* (K-Star) and Naive Bayes, respectively. For short text stream clustering, we named our method as T-GSC (A Two Level Graph Based Short Text Stream Clusterer). This method has two clustering phases and first one is based on embeddings generated from the short text documents where second one is generated from the graphs of clusters obtained in the first level. We also prepared a survey about the current methods in short text stream clustering and classified them with respect to their clustering approaches.