Structural Credit Assignment in Hierarchical Classification.
Joshua K. Jones, Ashok Kumar Goel · International Conference on Artificial Intelligence · 2007
Structured matching captures the bottom-up pattern of the hierarchical classification techniques used in many knowledge systems: subsets of features describing the state of the world are progressively aggregated into equivalence classes in an abstraction network, until the required decision is made at the root node. In this paper, we describe a supervised learning technique for automatic repair of abstraction hierarchies in structured matching. Note that the task of repairing an abstraction network involves the structural credit assignment problem. We describe the property of empirical determinability, and show that the design of an abstraction network according to this property enables structural credit assignment.