Hybrid intelligent approaches for business process sequential analysis.

Abstract

The quality of customer services is an important differentiator for service oriented com- panies like telecommunication providers. In order to deliver good customer service, the underlying processes within the operations of a company have to run smoothly and must be well controlled. It is of great importance to be able to predict if processes are likely to fail and to be aware of developing problems as early as possible. A failure in a customer service process typically results in a negative experience for a customer and companies are keen to avoid this from happening. Process performance prediction allows companies to pro-actively adapt with process execution in order to prevent process problems from affect- ing their customers. Process analytics is often compounded by a number of factors. Very often processes are only poorly documented because they have evolved over time together with the legacy IT systems that were used to implement them. The workflow data that is collected during process execution is high dimensional and can contain complex attributes and very diverse values. Since workflow data is sequential in nature, there are a number of data mining methods such as sequential pattern mining and probabilistic models that can be useful for predicting process transitions or process outcomes. None of these techniques alone can adequately cope with workflow data. The purpose of this thesis is to contribute a combination of methods that can analyse data from business process in execution in order to predict severe process incidents. In order to best exploit the sequential nature of the data we have used a number of sequential data mining approaches coupled with sequence alignment and a strategy for dealing with similar sequences. The methods have been applied to real process data from a large telecommunication provider and we have conducted a number of experiments demonstrating how to predict process steps and process outcomes. Finally, we show that the performance of the proposed models can be significantly improved if they are applied to individual clusters of workflow data rather than the complete set of process data.