- Rename: listrules -> ruleset ☐ Optimize Sole apply/predict di LeftMostLinearFormulas e su dataset ☐ apply_postprocess
☐ AbstractTrees interface for SoleModels ☐ AbstractInterpretationSet/AbstractLogiset are probably the same thing!!
☐ There is an interesting new pattern between DecisionTree and Branch!!! DecisionTree acts as a guard! Maybe it's better if it's not a type, but just a function that builds Branch's or Finals!
☐ The same goes for DecisionForests, which should be a sort of alias for Ensemble{DecisionTree}. Type to create: Ensemble, type to remove: DecisionForest.
☐ Conversion functions between models, for example: convert(::Type{DecisionList}, tree::DecisionList)
☐ Check formula2natlang and here, Exercise 2: https://github.com/aclai-lab/modal-symbolic-learning-course/blob/main/Day3-gesture-recognition.ipynb. It’s a (limited, experimental) function that translates formulas into English. We should add some further simplification for cases like "(∃ interval where (min[V5] ≥ 0.85)) and (∀ intervals (min[V1] < 0.43))"! When the formula is (∃ interval where (min[V5] ≥ X)), it means that, over the entire series (i.e., over the largest interval, min[V5] ≥ X).
☐ TODO translate MDT in the propositional case must remove the existential and global quantifiers. Maybe it already does…?
☐ remove nvariables logiset is wrong and dangerous
☐ remove nvariables for MultiLogiset: dangerous.
☐ Learn from FeaturedGraph's: https://yuehhua.github.io/GraphSignals.jl/stable/manual/featuredgraph/#Pass-FeaturedGraph-to-CUDA
☐ Switch the types in ScalarCondition, and require the parser to specify like ScalarCondition{FT}, or ScalarCondition, featvaltype = FT. One should be able to parse IA formulas from a string like this:
φ = parseformula("[L] ( min[V1] > 0.5 ∨ min[V2] ≤ 10 )", [box(IA_L)]; atom_parser = str->parsecondition(ScalarCondition, str))
φ = parseformula("[L] ( min[V1] > 0.5 ∨ min[V2] ≤ 10 )", SoleLogics.diamondsandboxes(IARelations); atom_parser = str->parsecondition(ScalarCondition, str))
☐ TODO or of multiformulas…?? Free SyntaxBranch…?
☐ datasets:
☐ TODO: note that when checking a formula like f1 > 3, everything gets memoized in the large table!!
add a parameter so that formulas meeting certain parameters are not cached normally in the large table, but instead, you just do a lookup for them.
☐ chained checking
TODO: apply_test_operator(test_operator, gamma, threshold(featcond)) with the result of compute_chained_threshold
optimize collateworlds with representatives aggr etc.
☐ non-uniform generalization
☐ AlphabetOfAny{C where C<:Condition{F where F<:Feature{P where P<:Atom}}}
☐ TODO generic relational/global structure that holds an indefinite number of relations/metaconditions
☐ scitype = Image/TimeSeries?
☐ Translate symbolic models into logical formulas. (ConstantModel("C_1") -> Atom(ConstantModel("C_1")))
Features: ☐ struct EnsembleModel (with parametrized aggregation) and DecisionForest ☐ PatchModel, a closed model providing a default consequent to an open model ✔ Add supporting_labels field (vector) @done(24-05-31 11:12) # Vector of labels of instances on which the rule is built (i.e., those not covered by previous rules), and in # the constant model info, include a supp_labels field with the labels of the instances that, among these, are covered by the rule. ☐ Add predicted_labels field (perhaps) # Labels predicted by the model, but initially the supporting_labels are sufficient # to calculate various metrics
Test: ✔ Testing parser outcome @done(24-05-31 11:12) ✔ Correct supporting_labels @done(24-05-31 11:12) ✔ Correct info (consequent/rule) @done(24-05-31 11:12) # Distribution of covered examples for consequent # Distribution of examples on which the rule was built ✔ Testing parser error @done(24-05-31 11:12)
Questions: ✔ Readmetrics for CN2 statistics @done(24-05-31 11:12)