A Comparative Study of Markov Network Structure Learning Methods Over Data Streams

Markov network is a widely used graphical representation of data in applications such as natural language and computational biology. This undirected graph consists of nodes and edges as attributes and its dependencies respectively. One major challenge in a learning task involving Markov network is to learn its structure, i.e. Attribute dependencies, from data. This has been the subject of various studies in the recent past, which uses heuristics to estimate dependencies from data. In this paper, we highlight the challenges of Markov network structure learning, and review existing methods addressing these challenges. In particular, we study the scalability of these heuristics over streaming data where data instances are assumed to occur continuously. Furthermore, we propose a new heuristic based on clustering of features, consisting of attribute dependencies, that can seamlessly update the model structure as new data arrive in a stream. This clustering technique effectively reduces search space and uses fewer number of features to generate a single model. Weight learning and inference is performed at the end of each data chunk consisting of data instances arriving within a fixed time frame. We empirically evaluate the proposed heuristic by comparing the CMLL score, on various datasets (both streaming and non-streaming), with other state-of-the-art methods.