Data Driven Profiling of Dynamic System Behavior Using Hidden Markov Model Based Combined Unsupervised and Supervised Classification

Abstract

Dynamic systems are often best characterized by a combination of static and temporal features, with the static features describing time-invariant properties of the system, and the temporal features capturing dynamic aspects of the system. Our goal is to construct context based temporal behavior models of dynamic systems using information from both types of features. Our dynamic system profiling framework consists of three main steps: (i) model generation, (ii) model idation, and (iii) model interpretation. Model generation step can be further decomposed into two components: (ia) temporal model generation, and (ib) context generation. Based on temporal feature values of the systems, temporal model generation step constructs K models to account for dynamic behavior patterns. We choose Hidden Markov Model(HMM)(R.abiner 1989) representation for temporal models. One important and desirable characteristic of HMM is that the hidden states of a HMM can effectively be used to model the set of potentially valid stages going through by a dynamic system and the directed probabilistic links between states be used to model its transition patterns among the set of stages. Our HMM clustering scheme tries to improve upon existing methods in two ways: First, existing HMM clustering systems assume fixed, pre-specified HMM topology. To obtain better fit models, we propose a dynamic and automatic HMM refinement procedure that interleaves with the clustering process and constructs HMMs of appropriate topologies for individual clusters. Bayesian model selection criteria(Chichering 8¢ Heckerman 1997) are employed in this process. Second, existing HMM clustering systems rely on predefined threshold values to determine number of clusters, i.e., the value of K, in the final partition. We take a model based approach(Cheeseman & Stutz 1996) Our clustering model is composed of clusters in the current partition and one hidden state that assigns cluster membership for each object. Given this clustering model structure, the number of clusters in the final partition is one that gives the highest model posterior probability. The K clusters derived from temporal model generICopyright @1999, American Association for Artificial