onlookerliu/pyscorecard

Python client for submitting PMML ScoreCard models and queries against them to OpenScoring

Note: I'm not currently working on this, if you're interested in picking it up, file an issue over on https://github.com/ncoghlan/admin-meta

pyscorecard

Python client for submitting PMML ScoreCard models and queries against them to OpenScoring

PMML ScoreCards: http://dmg.org/pmml/v4-2-1/Scorecard.html for additional background

OpenScoring REST API: https://github.com/jpmml/openscoring Demo instance: http://openscoring-ncoghlan.rhcloud.com/openscoring/ Demo git repo: https://github.com/ncoghlan/openscoring-openshift

PMML ScoreCard generation from JSON input

Command line invocation:

pyscorecard input_spec.json pmml_output_dir

Output PMML file names are generated based on a combination of "model_name" and "param_grid" entries as described below.

In the Python API, scorecard.pmml_scorecard generates PMML scorecard definitions from a JSON-compatible input mapping.

See examples/risk_example.json (input) and examples/risk_example.xml (output)

All ScoreCards produce a single predicted risk score and up to 3 reason codes:

• RiskScore
• ReasonCode1
• ReasonCode2
• ReasonCode3

Generated ScoreCards are also currently all hardcoded to use the "pointsAbove" reason code algorithm, the "min" baseline score algorithm, 0 as the initial score for the overall scorecard evaluation and 1 as the baseline score for each individual characteristic (this ensures that characteristics achieving a partial score of 0 are never reported as reason codes for the overall risk scoring).

The input format is a JSON mapping with the following fields:

• model_name: name of the model (also used as output filename prefix)

• param_grid: parameter definitions for use in characteristic predicates

  • key is the variable name that can be substituted into predicates
  • value is a mapping of output filename suffixes to substition values
  • when multiple grid parameters are defined, keys are lexically sorted when determining the combined output filename

• data_fields: sequence of field definitions for the DataDictionary and MiningSchema sections in the generated PMML Scorecard

  • name: used in both the DataField entry and the MiningField entry
  • dataType: used in the DataField entry
  • optype: used in the DataField entry to define handling of comparisons
  • values: permitted values for categorical and ordinal fields

• characteristics: sequence of definitions for the Characteristic section in the generated PMML Scorecard

  • name: data field used by this characteristic. Also used to derive the characteristic name as name + "Score" and the overall characteristic reason code as name + "RC"

  • attributes: sequence of attribute definitions used by the characteristic

    • reasonCode: specific reason code when this criterion is met
    • partialScore: contribution to the risk score when this criterion is met
    • predicate: predicate defining this criterion (see below for details)

Predicates can be defined as either a single string, or as a sequence of such strings. Each string predicate is of the form "OP value", with the data field named in the characteristic definition being the implied left hand side of the operation. Predicate sequences are implicitly and'ed together to define the overall criterion to be met for that attribute. Predicate values may start with $ to indicate a grid parameter - these will be substituted with the appropriate value for the scorecard currently being generated.

Permitted operations are == for data fields with the categorical optype, and ==, <, <=, >=, and > for data fields with the ordinal or continuous optype.