Regularizing conjunctive features for classification

We consider the feature-generation task wherein we are given a database with entities labeled as positive and negative examples, want to find feature queries that linearly separate two sets of examples. focus on conjunctive queries, explore problems: (a) deciding if separating exist (separability), (b) generating such when they exist. To restrict complexity generated classifiers, various ways regularizing them by limiting their dimension, number joins in generalized hypertreewidth (ghw). show separability problem is tractable for bounded ghw; yet, generation not because might be too large. So, third problem: classifying new without necessarily queries. Interestingly, case ghw can efficiently classify explicitly