This paper addresses the problem of target-directed exploration (TDE) in initially unknown and large-scale indoor environments. In such scenarios, the inference on an unknown space can improve the search performance under the assumption that the context of a particular space (i.e., the functional category of the space) is correlated with the existence of a target. The space inference is promising in that there is a strong statistical correlation between the semantic categories of indoor spaces and their adjacency because the spaces are designed to reflect universal human preferences. In this point of view, we propose a novel TDE scheme leveraging the semantic-spatial relations of an indoor floorplan dataset. Whereas existing works dealing with the data-driven space inferences consider only the one-to-one relation statistics of the spaces or utilize heuristic counting-based matching algorithms without building a trainable latent model, we propose the pattern cognitive Multivariate Bernoulli Distribution-based Graphical Space Inference Model (MBD-GSIM). MBD-GSIM efficiently captures the core contexts of the discrete semantic-spatial relations to predict an unknown space by using the latent Multivariate Bernoulli Distribution model. We also suggest utilizing the MBD-GSIM in a cost-utility based frontier exploration scheme for TDE problems. The proposed scheme is constructed in the Robot Operating System (ROS); its efficiency is investigated in the Gazebo simulation environment.