TD_GLM produces the following outputs:
- Model (Primary output): Contains the trained model with model statistics. The following model statistics are stored in the model:
- Loss Function
- MSE (Gaussian)
- Loglikelihood (Logistic)
- Number of Observations
- AIC
- BIC
- Number of Iterations
- Regularization
- Alpha (L1/L2/Elasticnet)
- Learning Rate (initial)
- Learning Rate (Final)
- Momentum
- Nesterov
- LocalSGD Iterations (Partition by Any only)
- [Optional for PARTITION BY ANY] MetaInformationTable (Secondary Output): Contains training progress information for each iteration. When the StepwiseDirection parameter is specified, the secondary output table contains information for the Stepwise Regression algorithm.
The model output schema for partition by any is as follows:
| Column | Data Type | Description |
|---|---|---|
| attribute | SMALLINT | Numeric index of predictor and model metrics. Intercept is specified using index 0, and the rest of the predictors take positive values. Model metrics take negative indices. |
| predictor | VARCHAR | Name of the predictor or model metric. |
| estimate | FLOAT | Predictor weights and numeric-based metric values. |
| value | VARCHAR | String-based metric values such as SQUARED_ERROR for LossFunction, L2 for Regularization, and so on. |
[Optional for PARTITION BY ANY] The MetaInformationTable output schema is as follows:
| Column | Data Type | Description |
|---|---|---|
| iteration | INTEGER | Iteration number. |
| num_rows | BIGINT | Total number of rows processed. |
| eta | FLOAT | Learning rate for the iteration. |
| loss | FLOAT | Loss in the iteration. |
| best_loss | FLOAT | Best loss until the specified iteration. |
| Step | INTEGER | [StepwiseDirection only] Iteration number (step number). |
| SubStep | INTEGER | [StepwiseDirection only] Feature number to be added or deleted. Non-feature numbers are given to different algorithm stages for sorting purposes. |
| Description | VARCHAR | [StepwiseDirection only] Description of the stages of each step. Added features are preceded by a plus sign, and deleted features by a minus sign. |
| Score | FLOAT | [StepwiseDirection only] Score of a given model tested in each substep, as well as the best score in each step and the best overall score, as indicated by the Description column. |
| Model | VARCHAR | [StepwiseDirection only] List of variable names that are contained in the model at each step. |
The output schema for partition by key is as follows:
| Column | Data Type | Description |
|---|---|---|
| partition_by_column | CHARACTER, VARCHAR,INTEGER, BIGINT, SMALLINT, BYTEINT | Data type is the same as the original column in the input table. |
| attribute | SMALLINIT | Numeric index of predictor and model metrics. Intercept is specified using index 0, and the rest of the predictors take positive values. Model metrics take negative indices. |
| predictor | VARCHAR | Name of the predictor or model metric. |
| estimate | FLOAT | Predictor weights and numeric-based metric values. |
| value | VARCHAR | String-based metric values, such as SQUARED_ERROR for LossFunction, L2 for Regularization, and so on. |