Judging the Performance of Cascading Models: A First Look

This paper explores different methods for interpreting the results of multiple, cascading machine learners, each of which performs a different task. For instance, the first learner may classify news as “sports,” the second learner may extract the people from the sports articles, and the third learner may classify the extracted people as belonging to a certain team. We present a framework for modeling such cascading learners as a directed acyclic graph, which allows us to construct three-way contingency tables on which we can perform various independence tests. These independence tests provide insight into how the various learners’ performance depend on their predecessor in the chain and/or the inputs themselves.