PROJECT TITLE :

Discovering Temporal Patterns for Event Sequence Clustering via Policy Mixture Model

ABSTRACT:

The Temporal Point Process, or TPP for short, is an expressive modeling tool that can be used to analyze the temporal pattern of event sequences. In TPP modeling, however, the process of identifying temporal patterns for the clustering of event sequences is only rarely studied. In order to solve this issue, we employ a technique known as reinforcement learning, which presupposes that the sequences that have been observed are the result of a combination of covert policies. The goal here is to learn each of the policy models while simultaneously clustering the sequences into the underlying policies based on the various temporal patterns they exhibit. The adaptability of our model is due to the following features: I all of the components, including the policy network for modeling the temporal point process, are networks; ii) to handle varying-length event sequences, we resort to inverse reinforcement learning by decomposing the observed sequence into states (RNN hidden embedding of history) and actions (time interval to next event) in order to learn a reward function, which helps to achieve better performance or increasing efficiency in comparison to existing models; and iii We use an algorithm called Expectation-Maximization, where in the E-step we try to estimate the cluster labels for each sequence and in the M-step we try to learn the respective policy. Extensive testing on both artificial and data taken from the real world demonstrates that our approach is competitive with the current best practices in the field.