Speeding Up Relational Data Mining by Learning to Estimate Candidate Hypothesis Scores
Frank DiMaio, JUDE W. SHAVLIK, W. Dayton · 2003
The motivation behind multi-relational data mining is knowledge discovery in relational databases containing multiple related tables. One difficulty relational data mining faces is managing intractably large hypothesis spaces. We attempt to overcome this difficulty by first sampling the hypothesis space. We generate a small set of hypotheses, uniformly sampled from the space of candidate hypotheses, and evaluate this set on actual data. These hypotheses and their corresponding evaluation scores serve as training data in learning an approximate hypothesis evaluator. We use this approximate evaluation to quickly rate potential hypotheses without needing to score them on actual data. We test our approximate clause evaluation algorithm using the popular Inductive Logic Programming (ILP) system Aleph. We use a neural network to approximate the hypothesis-evaluation function. The trained neural network replaces Aleph’s hypothesis evaluation on actual data, scoring potential rules in time independent of the number of examples. Our approximate evaluator can also be used in a heuristic search to help escape local maxima. We test the neural network's ability in learning the hypothesis-evaluation function on four benchmark ILP domains; the neural network is able to accurately approximate the hypothesis-evaluation function.