Oppia-ml-Extension.md

Oppia-ml Extension

Oppia-ml is developed and maintained by Oppia developers. It allows Oppia to train machine learning models so that Oppia can provide better answers to learners. Oppia-ml is developed in its own repo, separate from Oppia.

Communication with Oppia

Oppia-ml communicates with Oppia over the network for fetching job requests and storing result of job requests.

The following url handler is used for fetching job requests:

/ml/trainingjobhandler

• Request data: above handler requires following JSON data in the request:

  {
      ‘vm_id’: a unique ID assigned to VM.
      ‘signature’: digital signature for authenticity of VM.
  }
  

• Response data: above handler returns following JSON data in the response:

  {
      ‘job_id’: a unique ID using which Oppia can distinguish different jobs.
      ‘algorithm_id’: an ID using which Oppia-ml can identify the type of classifier to be trained.
      ‘training_data’: training data using which classifier will be trained.
  }
  

The following url handler is used for storing the result of job request:

/ml/trainedclassifierhandler

• Request data: Above handler requires following JSON data in the request:

  {
      ‘vm_id’: unique ID assigned to Oppia-ml instance.
      ‘message’: The dict that contains job_id and classifier_data, using which signature is generated.
      ‘signature’: digital signature for authenticity of VM.
  }
  

• Response data: It returns HTTP 200 in response if “classifier_data” has been saved successfully.

These URLs are defined in the oppia-ml/vmconf.py and oppia/feconf.py files. They must match the above ones, otherwise Oppia-ml will not be able to work with Oppia.

How to try your own algorithms.

The new oppia-ml project allows developers to try out their own ML algorithms for classification. The steps that need to be followed for using a new algorithm:

• Extend BaseClassifier in the oppia-ml repo and implement all its base functions for your chosen classifier.
• It is very important to add your algorithm to the Interaction-Classifier mapping. Without this step, your algorithm won't be functional.
• Implement a front-end service that will be called by the PredictionAlgorithmRegistryService. This service should have a predict() method which returns the predicted labels given user responses and classifier model.