matlab-classify-tabular-data

What it does

Classifies tabular data end-to-end in MATLAB: load a dataset, prepare and clean it, select promising classifiers, train them, and compare accuracies with cross-validation, holdout, or hyperparameter optimization plus statistical tests. Triggered when the user asks to classify tabular data, pick classifiers for a dataset, compare classifier accuracy, run cross-validation or a holdout evaluation, or find the best model with statistical uncertainty.

How it works

• Step 1 loads the data and determines whether there is a separate test set; it defines XTrain/YTrain/XTest/YTest or X/Y for use.
• Step 2 (Analyze and clean data) requires reading and following dataprep instructions from a referenced file, ensuring a flags struct and a preproc struct exist in the workspace as specified.
• Step 3 (Choose evaluation strategy) uses dataset size information to recommend either a 70/30 holdout split or 5-fold cross-validation, and asks the user to choose (holdout or cv). If the user chooses holdout, a holdout split is created via a script; if cv is chosen, set to cross-validation.
• Step 4 (Ask about interpretability) asks the user how important interpretability is, recording interpretability as 'high' or 'low' to steer model selection.
• Step 5 (Select and define model templates) requires reading an external reference to determine allowed classifiers, uses a helper to build model definitions, and may involve an imbalanced-prior workflow if applicable. It assembles modelDefs with fields like name, template, cvFitFcn, fitFcn, fnName, hyperparameters, innerLearner_fnName, innerLearner_hyperparameters. Multiclass ECOC recipes appear twice (OVO and OVA). The workflow trains every entry and uses CV paths or holdout paths accordingly; it does not perform raw fit calls like fitc* directly in the specified step.

The skill relies on external references and helper scripts to construct and train models, perform evaluations, and compare results. It enforces constraints on allowed priors and excludes certain neural-network-centric workflows.

When to use it

• You have a numeric matrix or a table of predictors with a categorical response and want to compare multiple classifiers.
• You need cross-validation, a holdout evaluation, or hyperparameter optimization for classifiers, plus statistical tests to judge significant accuracy differences.

What it can touch

• Inputs and outputs are manipulated through MATLAB code invoked via a MATLAB MCP server. The workflow uses predefined scripts and helper functions loaded from the skill path, and does not directly expose internal filenames in user-facing messages.

Caveats

• The skill does not support custom class priors or cost matrices; it relies on a built-in uniform-prior toggle for imbalanced data when applicable. If custom priors or costs are required, it instructs using dedicated fitc* calls directly.
• It requires reading external reference files for dataprep and classifier selections and relies on step-by-step prompts to gather user preferences (holdout vs cv, interpretability).
• It will not output raw code until the appropriate references are read and the user selections are made; it emphasizes messaging results to the user rather than internal file names.