WIP: new command: analyze - performs analysis of logical axioms

[cmungall]

This command will analyze all logical axioms in an ontology. So far only a single analysis is performed, testing the power of each axiom. See below for how this is performed.

This is an important metric, many novice ontology develops will add logical axioms without understanding the entailments, this provides a way of knowing which axioms are just added for decoration and do not provide any inferential benefit. There is some analogy to formal static analysis in computer programs and unreachable code but this is more nuanced and potentially worse.

This is what the current doc page for this looks like:

Analyze

Example:

robot analyze  --input logical_axioms.owl --output results/analysis_report.tsv

This performs an analysis of logical axioms in the ontology. Currently one type of analysis is performed, a power analysis of each axiom.

Each axiom is tested for its inferential power - i.e. how many additional axioms does it entail, and how many existing asserted axioms can only be entailed when it is used.

This is useful to know, as a low power rating means that you may be under-utilizing axioms and reasoning, or you may be overstating logical definitions.

Algorithm

Step 1: Removed Redundant Axioms

Any asserted axiom that can still be entailed is removed from the ontology

For each subClassOf axiom A = C SubClassOf D, we remove A from O. We then use the reasoner to test if D is an inferred superclass (direct or indirect of C).

If this is the case, the A is redundant and is removed from the ontology.

Step 2: Perform relaxation step

Additional SubClassOf axioms are added, according to the relax operation.

C EquivalentTo X1 and X2 and ... and Xn
==>
C SubClassOf X1
C SubClassOf X2
...
C SubClassOf Xn

The purpose of this step is to avoid treating the relaxed axioms as "new" inferences.

Step 3: Test individual axioms

For each logical axiom A we calculate:

Power = | E(O) - E(O-A) |

Where E returns all entailed direct axioms (currently just subClassOf) from an ontology.

Step 4: Aggregate statistics and report

The average power is calculated, as well as all axioms that share the maximum power.

Workflow

By design, this command works on the root ontology in an import chain. Use merge first if you need to merge an import chain into one ontology.

For example, if you wish to evaluate how much CL, CHEBI, and Uberon, etc contribute to classification in GO, then you can take either the go-edit file, or the release go-plus (both of which include full axioms, plus import modules), merge the results, then run analyze.

Currently aggregate stats will not be performed per-source, but you can easily do this from the output TSV using standard tools.

[cmungall]

If anyone wants to play with this there is a docker container cmungall/robot-analyze:

https://cloud.docker.com/u/cmungall/repository/docker/cmungall/robot-analyze

[cmungall]

Alternate names:

• inspect
• inspect-inferences
• analyze-inferences

Another way to think of this is as a mass semantic diff

[jamesaoverton]

I don't know the status of this, but I'm converting to Draft to prevent accidental merges and be a bit more organized.

[matentzn]

this should be merged with the measure command..

[jamesaoverton]

@matentzn says he wants this, so I've assigned it to him.

[jamesaoverton]

@matentzn Do you still want this, or is it all in measure now? I'd like to clean up these ancient PRs.

[matentzn]

Closing this and relegating the PR back to an issue, which I have pre-emptively closed as "not planned": https://github.com/ontodev/robot/issues/1167