Learning 'good quality' resource allocation from historical data
Effective and efficient delivery of services requires tasks to be allocated to appropriate and available set of resources. Much of the research in task allocation, model a system of tasks and resources and determine which tasks should be executed by which resources. These techniques when applied to service systems with human resources, model parameters that can be explicitly identified, such as worker efficiency, worker capability based on skills and expertise, authority derived from organizational positions and so on. However, in real-life workers have complex behaviors with varying efficiencies that are either unknown or are increasingly complex to model. Hence, resource allocation models that equate human performance to device or machine performance could provide inaccurate results. In this paper we use data from process execution logs to identify resource allocations that have resulted in an expected service quality, to guide future resource allocations. We evaluate data for a service system with 40 human workers for a period of 8 months. We build a learning model using Support Vector Machine (SVM), that predicts the quality of service for specific allocation of tasks to workers. The SVM based classifier is able to predict service quality with 80% accuracy. Further, a latent discriminant classifier, uses the number of tasks pending in a worker's queue as a key predictor, to predict the likelihood of allocating a new incoming request to the worker. A simulation model that incorporates the dispatching policy based on worker's pending tasks shows an improved service quality and utilization of service workers.
Please refer to publisher version or contact your library.