It has been very difficult to gather full data sets from institutions with which to develop, train, and test learning analytics models. We hope that more models will be developed in partnership with institutions, or independently by institutions using Moodle. To help facilitate this, we could provide a service that automates sharing models and model training data sets. This service might become part of MoodleNet, but would require work within Moodle to facilitate ease of use.

Moodle would need the following features in Core:

1. Opt-in trigger in site registration to submit data on what models are in use on the system, their accuracy, and their usage rate (i.e. how many users are receiving insights and what proportion of those insights are being viewed and/or acted on).  Some metadata about the site should also be included, i.e. from MDL-57900. Note that some of this data is not currently collected within Moodle sites, but there are proposals to do so, e.g. MDL-62192, MDL-62302.
2. Administration tool to submit a model definition in a reusable form (e.g. PMML) to the Moodle-supplied service - see MDL-60944
3. Administration tool to submit non-identifying model data in csv form to the Moodle-supplied service, identified by a hash code or other mechanisim to tie the data to the model used to generate it - this exists as a manual export capability now, but submission needs to be very easy (even with an automatic option) to encourage sites to submit.
4. Administration tool to select, download and import a model definition, as generated in #2, based on information provided about the model accuracy, testing status, and usage (number of sites, number of users, type of site, as collected in #1)
5. Administration tool to select, download and import model training data, as generated in #3, based on information provided about the accuracy of the model when trained on that data set across multiple sites, types of sites, etc.

I'm sure this proposal needs to be refined further.