matlab-use-machine-learning-apps

Open Classification Learner or Regression Learner with your data table, create models from the gallery presets, and train them together. Compare results using the Results Table and Compare Results plot, which show validation metrics like accuracy, RMSE, prediction speed, and model size side by side.

Use the export to Simulink feature, available in MATLAB R2024a and later, to generate a Simulink block from the trained model. Models trained with categorical predictors cannot be exported to Simulink, so encode categoricals as numeric values before loading data.

Yes, optimizable model presets run Bayesian hyperparameter optimization with 30 iterations by default. You can switch to grid or random search, set training time limits, or export to Experiment Manager for finer control over search ranges and constraints.

Yes, Local Shapley and LIME plots are available from R2023b, and Shapley importance, summary, and dependence plots from R2024b. Shapley computation can be slow on large datasets or complex models, so adjust query points and observation samples to control cost.

Avoid these apps for command-line ML workflows where fitc or fitr functions are more direct, and for deep learning on non-tabular data like images, sequences, or text, which requires Deep Learning Toolbox. Data preprocessing and cleaning must also be done before importing data into the app.

Preprocessing that uses aggregate dataset statistics, such as mean imputation or z-score outlier removal done before import, leaks information from validation folds into training. The app does not apply these operations inside its CV folds, so perform such preprocessing within a fold-aware pipeline instead.