[
https://issues.apache.org/jira/browse/ATLAS-3570?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Barbara Eckman reassigned ATLAS-3570:
-------------------------------------
Assignee: Barbara Eckman
> Atlas typedefs for Machine Learning Models, Feature Sets, and Feature
> Engineering Engines
> -----------------------------------------------------------------------------------------
>
> Key: ATLAS-3570
> URL: https://issues.apache.org/jira/browse/ATLAS-3570
> Project: Atlas
> Issue Type: New Feature
> Reporter: Barbara Eckman
> Assignee: Barbara Eckman
> Priority: Major
> Attachments: MLModel_typedefs.tar
>
>
> Currently the base types in Atlas do not include Machine Learning (ML) Model
> tables. It would be nice to add typedefs for them, so they could be part of
> enterprise discovery and versioning.
> ENTITIES COULD INCLUDE:
> MLModel (overview info), with attributes:
> * uniqueId
> * version
> * businessUseCase
> * modelFramework (eg scikit-learn)
> * modelTypes (eg random forest regressor)
> * modelClass (eg random forest (bagging + decision trees))
> * isEnsemble boolean
> * outcomeTypeDescription (eg single float)
> * **dataScienceOwnerEmail
> * githubRepoURL where the model code is founc
> * modelDeploymentDate
> * populationScored (eg in Comcast, residential or business customers)
> * accuracyMeasures
> MLModelExecution, with attributes:
> * exampleInputDatasetURL (URL where a sample input dataset can be found)
> * outputTargetDatasetURLs
> * opsOwnerEmail
> * executionEndpointURL
> * dockerContainerURL
> * MLFlowPointerURL
> * executionNotebookURL (eg Databricks, Jupyter)
> MLModelTraining, with attributes:
> * hyperParameters
> * trainingDatasetURLs
> * trainingNotebookURL (eg Databricks, Jupyter)
> FeatureSet (a set of features prepared as input to an ML model), with
> attributes:
> * version
> * locationURL
> FeatureEngineeringEngine (the engine that generates the feature set for an ML
> model), with attributes:
> * version
> * ownerEmail
> * inputSourceURL
> * processingEngineInfoURL (docs on the processing engine)
> * githubRepoURL
> * outputTargetURL
> RELATIONSHIPS could include:
> * model to execution
> * model to training
> * model execution to example input dataset (eg kafka topic)
> * model execution to output target dataset (eg S3 prefix or object)
> * model execution to input schema
> * model execution to output schema
> * model execution to input feature set objects
> * training to input training dataset objects
> * training to input training dataset schema
> * feature engineering engine to output feature set object
> * feature engineering engine to input source dataset (eg kafka topic)
> * feature engineering engine to input source dataset's schema
> * feature engineering engine to output target dataset (eg S3 object)
> * feature set object to its schema
> ENUMs could include:
> * MLModel_type (eg logistic regression, random_forest_regression)
> PROCESSES related to MLModels could include:
> * MLPipelineDependencyEdge (dependency between two models in the ML pipeline)
> ** inputs and outputs are both MLModels
> * MLModelEvolutionEdge (lineage between 2 versions of an ML model)
> ** inputs and outputs are both MLModels
> ** only attribute is an array of strings representing changes made from one
> version to the other. this could be made more structured as we discover how
> it is used.
>
--
This message was sent by Atlassian Jira
(v8.3.4#803005)