Unsupervised clustering of service performance behaviors

We propose in this paper a novel approach for unsupervised clustering of services' behaviors. These behaviors are modeled as multivariate time series that capture the evaluation of several service quality attributes for a period of time. The importance weights of quality attributes are derived based on the Shannon's entropy concept and the service data is flattened in a format that is convenient for clustering. The flattening process spans over a time oriented aggregation transformation, which leverages Haar reduction. The reduction is modeled as a maximization of an objective function. The absence of ground truth is tackled by performing a set of tests to determine the best number of clusters and clustering algorithms. Extensive experiments were conducted to validate the proposed unsupervised clustering approach.