Prediction Service User Manual with Alibi Anchors

Introduction

This service will let you to generate the prodictions explained for sigle observations using a previously trained model with AnchorTabular method from alibi-explain library. The service is though up to operate only in combination with a compatible traning service of the Data Analytics System platform, producing models in a supported format.

Service Features

Make a prediction for one or more input rows.
Generate loca explanations in form of anchors (rules).
Support the customization of Alibi method explain() parameters.
Manage the model recovery, explainer recovery and input strategies recovery.

Parameters

--direct_args_to_explainer_function: JSON dictonary with methos parameters .explain() of the explainer. For complete details on accepted parameters, please check the official documentation: AnchorTabular.explain

Input and Output

Input

Input Model: it must be a trained model with a Data Analytics System traning service based on AnchorTabular. The format dictate the model contains also the explainer and the informazions for the input.
Dataset: it can contain one or more observations (rows), as long as it is provided as "streaming" dataset type in JSON format dictionary or a list of dictionaries.

Output

The output message contains the origin input added with predictions field including:
"name" (target class),
"value" (model prediction),
"explanation" (anchoref rule in text format).

Use Case Sample

Input Example (one row):

{
  "age": 45,
  "job": "technician",
  "marital": "married",
  "education": "secondary",
  "balance": 1234
}

Parameters:

{
  "direct_args_to_explainer_function": {
    "threshold": 0.8,
    "beam_size": 10
  }
}

Output:

{
  "age": 45,
  "job": "technician",
  "marital": "married",
  "education": "secondary",
  "balance": 1234,
  "predictions": [
    {
      "name": "predicted_label",
      "value": 1,
      "explanation": "IF job = technician AND education = secondary THEN prediction = 1"
    }
  ]
}

Important Notes

This service is not indipendent: it needs models build with alibi-explain-anchors training service available on Data Analytics System platform.
If model requested features are missing, they will be added and input as automatically.
The input dataset have to be streaming type.