Pattern detection in unstructured data: An experience for a virtualized IT infrastructure

Data-agnostic management of today's virtualized and cloud IT infrastructures motivates statistical inference from unstructured or semi-structured data. We introduce a universal approach to the determination of statistically relevant patterns in unstructured data, and then showcase its application to log data of a Virtual Center (VMware's virtualization management software). The premise of this study is that the unstructured data can be converted into events, where an event is defined by time, source, and a series of attributes. Every event can have any number of attributes but all must have a time stamp and optionally a source of origination (be it a server, a location, a business process, etc.) The statistical relevance of the data can then be made clear via determining the joint and prior probabilities of events using a discrete probability computation. From this we construct a Directed Virtual Graph with nodes representing events and the branches representing the conditional probabilities between two events. Employing information-theoretic measures the graphs are reduced to a subset of relevant nodes and connections. Moreover, the information contained in the unstructured data set is extracted from these graphs by detecting particular patterns of interest.