The machine learning meta-model with synthetic data (useful for MLOps/feature store), is independent of machine learning solutions (definition in json, data in csv/parquet).
The meta-model is suitable for:
- compare capabilities and functions of machine learning solutions (as part of RFP/X and SWOT analysis)
- independent test new versions of machine learning solutions (with aim to keep quality in time)
- unit, sanity, smoke, system, integration, regression, function, acceptance, performance, shadow, ... tests
- external test coverage (in case, that internal test coverage is not available or weak)
- etc.
Note: You can see real usage of this meta-model in e.g. project qgate-sln-mlrun for testing MLRun/Iguazio solution.
The solution contains this simple structure:
- 00-high-level
- The high-level view to the meta-model for better understanding
- 01-model
- The definition contains 01-projects, 02-feature sets, 03-feature vectors, 04-pipelines, 05-ml models, etc.
- 02-data
- The synthetic data for meta-model in CSV/GZ and parquet formats for party, contact, relation, account, transaction, event, communication, etc.
- You can also generate your own dataset with requested size (see samples
./02-data/03-size-10k.sh,./02-data/04-size-50k.sh, etc. and descriptionpython main.py generate --help)
- 03-test
- The information for test simplification e.g. feature vector vs on/off-line data, test/data hints, etc.
Addition details, see structure and see rules
The supported sources/targets for realization (✅ done, ✔ in-progress, ❌ planned), see
the definition /spec/targets/ in projects (see specification in JSON files):
- ✅ Redis, ✅ MySQL, ✅ Postgres, ✅ Kafka
- ✅ Pandas, ✅ Parquet, ✅ CSV
The object relations for key objects in meta-model, plus splitting these objects in packages (01-model/01-project, 01-model/02-feature-set, 02-data, etc.).
The basic feature sets and relations between them.
The derived feature sets and relations between them.
