A Stochastic Model-Based Approach to Online Event Prediction and Response Scheduling

In a variety of contexts, time-stamped and typed event logs enable the construction of a stochastic model capturing the sequencing and timing of observable discrete events. This model can serve various objectives including: diagnosis of the current state; prediction of its evolution over time; scheduling of response actions. We propose a technique that supports online scheduling of actions based on a prediction of the model state evolution: the model is derived automatically by customizing the general structure of a semi-Markov process so as to fit the statistics of observed logs; the prediction is updated whenever any observable event changes the current state estimation; the (continuous) time point of the next scheduled action is decided according to policies based on the estimated distribution of the time to given critical states. Experimental results are reported to characterize the applicability of the approach with respect to general properties of the statistics of observable events and with respect to a specific reference dataset from the context of Ambient Assisted Living.