Markov networks are extensively used to model complex sequential, spatial, and relational interactions in a wide range of fields. By learning the structure of independences of a domain, more accurate joint probability distributions can be obtained for inference tasks or, more directly, for interpreting the most significant relations among the variables. However, the performance of current available methods for learning the structure is heavily dependent on the choice of two factors: the structure representation, and the approach for learning such representation. This work follows the probabilistic maximum-a-posteriori approach for learning undirected graph structures, which has gained interest recently. Thus, the Blankets Joint Posterior score is designed for computing the posterior probability of structures given data. In particular, the score proposed can improve the learning process when the solution structure is irregular (that is, when there exists an imbalance in the number of edges over the nodes), which is a property present in many real-world networks. The approximation proposed computes the joint posterior distribution from the collection of Markov blankets of the structure. Essentially, a series of conditional distributions are calculated by using, information about other Markov blankets in the network as evidence. Our experimental results demonstrate that the proposed score has better sample complexity for learning irregular structures, when compared to state-of-the-art scores. By considering optimization with greedy hill-climbing search, we prove for several study cases that our score identifies structures with fewer errors than competitors.
from cs.AI updates on arXiv.org http://ift.tt/2aVxFar
via IFTTT
No comments:
Post a Comment