Constructor: A System for the Induction of Probabilistic Models

The probabilistic network technology is a knowledgebased technique which focuses on reasoning under uncertainty. Because of its well defined semantics and solid theoretical foundations, the technology is finding increasing application in fields such as medical diagnosis, machine vision, military situation assessment , petroleum exploration, and information retrieval. However, like other knowledge-based techniques, acquiring the qualitative and quantitative information needed to build these networks can be highly labor-intensive. CONSTRUCTQR integrates techniques and concepts from probabilistic networks, artificial intelligence, and statistics in order to induce Markov networks (i.e., undirected probabilistic networks). The resulting networks are useful both qualitatively for concept organization and quantitatively for the assessment of new data. The primary goal of CONSTRUCTOR is to find qualitative structure from data. CONSTRUCTOR finds structure by first, modeling each feature in a data set as a node in a Markov network and secondly, by finding the neighbors of each node in the network. In Markov networks, the neighbors of a node have the property of being the smallest set of nodes which “shield” the node from being affected by other nodes in the graph. This property is used in a heuristic search to identify each node’s neighbors. The traditional x2 test for independence is used to test if a set of nodes “shield” another node. Cross-validation is used to estimate the quality of alternative structures.