Service selection based on Reputation using Statistical Measures (T-0727) (MFN 6194)

Abstract

Service oriented architecture depends on service selection based on user requirements. Historically service selection is orchestration based on scenarios. A single organization may not be able to build and host all the required services in a competitive market. With the increase of demand and availability of multiple services from specialist providers for a single task, there is need for dynamic service selection. This research identifies types of Quality of Service parameters, their impact on the overall service performance and a statistical model to help in selection of service using their historical performance measurements. Chebyshev inequality is used to find outliers in the data and the impact of removing outliers from data is discussed along with results. Gumbel and normal distributions are used to calculate reputation score. Existing techniques of reputation that depend on sum or average are heavily influenced by the lower or upper rare events. Belief, fuzzy, data mining and flow models are complex and in most of the cases require pre-and post-analysis along with complex calculation to form a decision. Probability distributions doesn’t require prior knowledge and works with small data set comparatively. In result and analysis, services were picked as defined by their behavior rather than their rare events. This also enables to analyze any set of data according to time or location references to understand the behavior without involving all the historic data. Prior to using distributions, Chebyshev inequality helped removal of up to 5% of outliers to further decrease the chances of rare events to influence the outcome.