[Demirrr]

Problem

With use_inverse=True, TDL added the feature ∀ r⁻.⊤ next to ∃ r⁻.⊤ for every individual that has a predecessor. With the filler ⊤, ∀ r⁻.⊤ is a tautology (the reasoner returns every individual), but its column was set only for individuals with a predecessor. So:

• the training matrix disagreed with the reasoner: 4 of 352 columns and 123 of 26,752 cells (0.46%) on the Family Uncle problem, all of the form ∀ r⁻.⊤ (hasChild, hasParent, hasSibling, married), always False in the matrix where the reasoner says True;
• the column was an exact copy of ∃ r⁻.⊤ and, because columns are sorted by DL string and ∀ sorts before ∃, it won ties against it, so a split on it was read back as ⊤ or ⊥.

Found while benchmarking a forest variant of TDL (research write-up in #672, follow-up 9).

Change

∀ r⁻.⊤ is no longer created (ontolearn/learners/tree_learner.py). New regression test TestTrainingMatrixSemantics in tests/test_tdl.py asserts that every cell of the training matrix agrees with the reasoner; it fails on develop and passes here.

Effect on learned hypotheses (behaviour change)

Hypotheses can change. 5-fold CV, held-out F1 with positive_paths, 18 Family problems (single CV run):

problem	before	after
Cousin	0.8223	0.7768
Grandgrandson	0.8985	0.8818
PersonWithASibling	0.9302	1.0000
Uncle	0.8929	0.9060

The other 14 problems are unchanged; the mean over the 18 problems goes from 0.9647 to 0.9658. Two problems get worse, so this is a correctness fix, not an accuracy improvement. tests/test_tdl_hypothesis_extraction.py pins CV values; the Cousin and Uncle full-tree values (and its docstring table) are updated to the new numbers. CHANGELOG.md has an entry.

Not covered

• Mutagenesis / Carcinogenesis were not re-measured, so the effect on large KBs is unknown.
• Only TDL-related tests were run: tests/test_tdl*.py (109 passed, 2 skipped) and tests/test_example_concept_learning_evaluation.py (passed). The full suite was not run. tests/test_triplestore.py needs a Fuseki server and was not run.

🤖 Generated with Claude Code