ML Model Tester: A Hands-On Machine Learning Lab

The ML Model Tester under AI Tools is a hands-on lab for machine learning: pick a model type, an algorithm, and a dataset, tune the hyperparameters, and train a real model right in the browser — then read genuine metrics, a confusion matrix, feature importances, a learning curve, and an ROC curve. Every number comes from an actual scikit-learn fit running on the server, so the results behave exactly like real machine learning, because they are.

Setting Up a Run

The page is organized as a workbench. On the left, choose a model type (classification, regression, clustering, and more), an algorithm (each with a one-line description of its strengths), and a dataset — including standard benchmarks like Wine Quality, Iris, and Digits, plus a DP Sound Systems supply-chain set. Hyperparameters — train/test split, cross-validation folds, number of estimators, max depth — sit just below.

The Performance Metrics panel adapts to the task. Classification shows accuracy, precision, recall, F1, and AUC-ROC, with cross-validation fold scores below. Regression swaps in R², RMSE, MAE, and MSE. Clustering reports silhouette score and inertia. The Confusion Matrix renders at the true class count — three classes for wine, ten for digits — with correct predictions on the diagonal.

Reading the Model

The Feature Importance chart names the actual columns that drove the model's decisions — not generic placeholders — so the output is interpretable against the domain. The Learning Curve plots training versus validation performance as data grows, and calls out the gap between them: a small gap means a healthy fit, a large one signals overfitting. That single diagnostic is one of the most useful things a modeler can see.

Regression on Business Data

Switching to regression on the DP Sound supply-chain dataset shows the tester on domain data. A Random Forest Regressor predicting demand reaches an R² of 95%, and the feature importances put demand at the top — the model correctly identifies the dominant driver in the data.

Experiments Worth Running

Because the training is real, the tester rewards experimentation — the results respond to what you change, and change in ways that teach:

1. Tune a hyperparameter. Set max depth from 10 to 1 on the Wine tree and watch accuracy fall sharply — a shallow tree cannot separate three cultivars. Raise it back and accuracy returns.
2. Compare algorithms on one dataset. Run Logistic Regression, then Random Forest, then SVM on the same data and compare their confusion matrices and AUCs.
3. Watch the learning curve. A widening train-validation gap is overfitting; try more cross-validation folds or a simpler model and see it narrow.
4. Read the importances against the domain. Flavanoids for wine, petal width for iris, demand for a demand-driven target — the model surfaces what actually matters.
5. Get AI Analysis. After a run, the built-in insights button has the local model interpret the metrics in plain language.

Try It

Open AI Tools → ML Tester, start with Wine Quality and a Decision Tree, and train. Then change one thing at a time — the algorithm, the depth, the dataset — and watch every panel respond. It is a fast, safe way to build intuition for how models actually behave, from confusion matrices to overfitting, without writing a line of code.

Comments (0)