TRACDS: Temporal Relationship Among Clusters for Data Streams

In this paper we propose a new extension to clustering data streams based on the Temporal Relationship Among Clusters for Data Streams (TRACDS). This is not a new clustering algorithm, but rather a way to capture the temporal relationships among clusters that is inherent in the ordering of observations in the data stream. We propose to capture this ordering relationship among the clusters by overlaying clusters created by any data stream clustering algorithm with a Markov Chain (MC). The states in the Markov Chain represent the clusters and the transitions are the relationships between clusters. For the TRACDS framework we identify the basic clustering operations used by any state-of-the-art data stream clustering algorithms and show how these operations can be mirrored on the MC using a set of TRACDS operations. This approach makes TRACDS general enough to be built on top of any clustering algorithm. We evaluate the improvement of TRACDS over just data stream clustering for outlier detection using several synthetic data sets to simulate different types of temporal structure as well as different levels of data quality (i.e., data points arrive outof-order). and Retrieval Clustering