Call minutes

Minutes from the ML Schema call June 20, 2016 Decisions: We have agreed that it would be good to try to finalize the draft documentation before most of us are gone for holidays. We need to make clear in the documentation: What it is for? (maybe link to the goals of our group and use cases), Why? (again look into the goals of the group such as to align existing schemas, to avoid proliferation of very similar resources), And how can ML Schema be used? (Provide examples and how to link to other resources). Some issues identified to achieve this:

1. Documentation - Introduction:

• to add the text on the motivation for ML Schema, e.g. to align existing ontologies and schemas (cite them in the references?), that is why we propose only highlevel, lightweight model
• to also motivate by the need for reproducible research
• to add something on the Audience („This document is mainly addressed to ML researchers / practisioners,..”, „for them to accomplish specific goals...” etc.)

2. Documentation - linking to other resources:

• to describe (in the section after introduction of the core model), that this proposed schema is complaint with other resources and can be used together with other ontologies and resources to provide more detailed information, e.g. with: DM/ML ontologies and schemas, software ontologies, PROV, Investigation-Study-Assay, Datatype ontologies etc.

3. Documentation - other:

• to cite OpenML and say that the example is derived from OpenML
• add to the current example the information on the task type (that :task29 is of type ClassificationTask). Add this to the text and also add this to the example code in turtle (with another namespace outside ML Schema core).
• explain what this schema in NOT meant for? (e.g., it is not meant to be a comprehensive ontology of ML which is going to replace existing models - those are already quite comprehensive, having various goals and are not going to be replaced)
• Incorporate the information on the envisaged use cases? They are listed on the Wiki of the group.

The way we are going to proceed: @Diego will be working on the points 1) and 2) until the end of the week (or next Monday the latest) and later on he will pass it to the next person who will pass it to @Larisa on July 1st.

Other news:

• @Tommaso and @Joaquin are working hard on openML2rdf code which is close to being complete.

Minutes from the ML Schema call June 6, 2016

Decisions:

• Documentation: extend the text, finish UML-based figure, add another figure with an example concerning instances; Should we mention use cases (OpenML)?
• Properties: better not to restrict domain and ranges (or not too much) in order to not prevent interoperability and re-use (see for instance the recent talk of Michael Uschold at Know@LOD2016)
• To release quickly a draft document and a call for feedback to the W3C mailing list and some other mailing lists (semantic web, machine learning)
• Create Twitter account? (also for disseminating the news on draft model)
• The importance of tools; To check openML2rdf code (@Joaquin); Converter from RDF to JSON-LD and back available from OpenML
• another use case (@Pance?)

Minutes from the ML Schema call May 23, 2016

Decisions:

• to add a link between Run and HyperParameterSetting to reflect that HyperParameterSetting is an input to Run (OpenML follows similar modeling) --> Issue #19
• to make optional that Task is "definedOn some evaluation specification"; instead reverse this relation and make a link from EvalutionSpecification to Task (e.g.: isAssociatedWith) --> Issue #20
• feature & data (issue #15): do no more than there is now and leave it as it is now
• to make optional that Run "has output some model evaluation" --> Issue #21
• ML software (Issue #18): yes, it is good to incorporate it; at the end we can simply add a class Software and use an external vocabulary / ontology (such as software ontology SWO or https://en.wikipedia.org/wiki/DOAP) to model whatever is needed such as a library name (scikit-learn, RapidMiner, ...), a version, a project name etc.
• use gh-page so it is more convenient to view the documentation --> DONE
• to decide what to do further with the diagram
• to create an issue on whether (and how) we should model relations' domains and ranges --> Issue #22