Creating Models

📄️ Creating Models

The simSearch, simRecommend, simClassify, and simCluster model types and how to start creating one from the Models page.

Why simClassify and simClassify+ require parameter tuning, unlike simSearch, and how classification use cases affect it.

Hyperparameter tuning finds the best Model Specification values, guided by the metric that matters most for the use case.

Grid Experiments train and evaluate many models across combinations of hyperparameter values to find the best-performing set.

Auto Tune searches a large hyperparameter grid automatically, often finding good results after evaluating only a third of it.

Exhaustive Grid Search evaluates every parameter combination in the grid, rather than skipping combinations as Auto Tune does.

ML Studio evaluates models using either Fold or Date Split experiments, chosen based on the training data set.

Fold Experiment Mode evaluates any model using a specified number of folds (N), independent of the data types involved.

Date Split Experiment Mode trains and tests a model by dividing a date-columned data set at two chosen split dates.

How to read the Grid Table and Grid Result Table & Model Creation tabs on the Grid Results page.

The Grid Analysis ROC Curve report compares AUC and True/False Positive Rate behavior across models with similar AUC values.

How to read a multinomial fold experiment's results table and use the Visualization option to inspect detailed metrics.

Domain Optimization lets simClassify+ optimize for a REAL-valued domain attribute, such as dollar amount, rather than count alone.

Weighted Recall measures the percentage of a chosen REAL value (e.g. fraud dollars) a model catches, not just record count.

Thresholding splits a simClassify or simClassify+ confidence value into a true/false prediction at a chosen cutoff.

Model Calibration trains an Isotonic Regression model to adjust a simClassify+ model's prediction probabilities.