All data that Predictable generates as part of the training and scoring process.

Model Train Subset

The training population of customers used to train a specific client model.

COLUMN NAME TYPE DEFINITION
CUSTOMER_ID VARCHAR unique customer identifier

Model Summary

Summary statistics of trained models. Area Under Curve (AUC) = overall summary statistic.

COLUMN NAME TYPE DEFINITION
MODEL_VERSION FLOAT version that model was trained on
TIMESTAMP NUMBER unix timestamp in seconds of when model was trained
TRAIN_ROC_AUC FLOAT summary statistic that evaluates overall predictive power on training set
TEST_ROC_AUC FLOAT summary statistic that evaluates overall predictive power on test set
TRUE_NEGATIVES FLOAT percentage of accurate negative predictions
FALSE_NEGATIVES FLOAT percentage of inaccurate negative predictions
TRUE POSITIVES FLOAT percentage of accurate positive predictions
FALSE POSITIVES FLOAT percentage of inaccurate positive predictions
NUMBER_OF_FEATURES INT total number of features processed in the model

Feature Importance

Lists all features in the model. A higher feature_value indicates a larger impact on prediction.

COLUMN NAME TYPE DEFINITION
MODEL_VERSION FLOAT version that model was trained on
TIMESTAMP NUMBER unix timestamp in seconds of when model was trained
FEATURE_NAMES FLOAT name of the feature (variable) included in the model
FEATURE_VALUES FLOAT normalized score (0-100) of impact feature had on predicted outcome

Product Matrix

Output of the product recommendation model.

COLUMN NAME TYPE DEFINITION
MODEL_VERSION FLOAT version that product recommendation model was trained on
TIMESTAMP NUMBER unix timestamp in seconds of when model was trained
PROD_OBS VARCHAR product that was purchased
PROD_REC VARCHAR product that is recommended
SCORE FLOAT metric that indicates the strength of the relationship – higher the better

Model Scores

Output of the churn, next purchase, and propensity models.

COLUMN NAME TYPE DEFINITION
CUSTOMER_ID STRING the unique customer identifier, the key to join back to customer tables.
MODEL_VERSION FLOAT software version the model was trained on for auditing purposes.
TIMESTAMP NUMBER unix timestamp in seconds of when scoring occured​. Used to get latest score for a given customer, and to filter out older scores for customers that have not been rescored.
SCORE NUMBER 0-100 normalized output of scoring.

Audiences

Predictable returns audiences in addition to scores. The audiences manifest as views, and contain a single column:

COLUMN NAME TYPE DEFINITION
CUSTOMER_ID STRING the unique customer identifier, the key to join back to customer tables.

The audiences are of two categories:

  • Standard audiences. These are high/med/low groupings for each model ouput, and are provided as a convenient pre-defined way to work with scores.
  • Smart audiences. These are based on customer attributes, scores, or some combination thereof.