Machine Learning Studio

Regression - predict a number

Predict weekly demand from price. Least-squares fits a curve; raise the degree for curvature. R² is the share of variation explained. A scanning line shows the live prediction.

Swipe the chart sideways to see all of it.

Noise Degree Predict at price

–

R² (fit quality)

–

predicted demand at selected price

–

fitted model

Supervised · numeric target. The line minimises the sum of squared vertical errors. Degree-1 assumes a straight relationship; higher degrees bend to fit - but too much degree chases noise (overfitting), when R² barely improves while the curve wiggles.

Logistic Regression - predict a category

Should an order be expedited? Logistic regression draws a straight boundary and, unlike a hard split, gives a probability - the shading is P(expedite). Watch it train by gradient descent.

Swipe the chart sideways to see all of it.

–

accuracy

0

training steps

Supervised · categorical target. It fits weights so a straight line best separates the classes, then squashes distance-from-line through a sigmoid into a 0–1 probability. Compared with a decision tree, the boundary is a smooth line, not rectangles - and every prediction carries a confidence.

Forecasting - predict future values

Three years of monthly demand with trend and seasonality. Holt-Winters learns level, trend and season, then projects ahead with an uncertainty band. Compare it to a naïve and moving-average baseline.

Swipe the chart sideways to see all of it.

Horizon

–

back-test error (MAPE)

–

next-month forecast

Supervised · time-ordered target. Naïve repeats the last value; a moving average smooths; Holt-Winters models the upward trend and the repeating seasonal shape, so its forecast keeps rising and wobbling with the season. The shaded band widens with the horizon - uncertainty grows the further out you look.

Clustering - find groups

Group stores by size and demand with k-means. Pick k and run: points join their nearest centroid, centroids glide to the group mean, and inertia falls until it settles.

Swipe the chart sideways to see all of it.

Clusters k

0

iterations

–

inertia (within-cluster spread)

Unsupervised · no labels. k-means minimises inertia - total squared distance from points to their centroid. Watch inertia fall and flatten as it converges. Different random starts can land in different groupings, which is why it's run several times.

Decision Tree - explainable rules

Flag which orders to expedite from weight and distance. The tree makes axis-aligned splits that cut impurity most - the coloured regions are the rules it learns.

Swipe the chart sideways to see all of it.

Max depth

–

training accuracy

–

decision regions (leaves)

Supervised · categorical target. Each split asks one yes/no question on one feature (e.g. "distance > 60?"). Shallow trees are readable but blocky; deep trees fit training data almost perfectly - often too well. The boundary is always made of rectangles.

Random Forest / Gradient Boosting - ensembles

Same expedite problem. One shallow tree is blocky; averaging many trees, each on a bootstrap sample, smooths the boundary and lifts accuracy. Slide the number of trees up.

Swipe the chart sideways to see all of it.

Trees Boosting view

–

accuracy

1

trees in the ensemble

Many weak learners → one strong one. A random forest grows many trees on random bootstrap samples and votes - variance cancels, so the boundary smooths and generalises. Gradient boosting adds trees in sequence, each fixing the last's errors - it climbs accuracy faster but overfits more easily. Toggle to compare the curve.

Machine Learning Studio

Regression - predict a number

Logistic Regression - predict a category

Forecasting - predict future values

Clustering - find groups

Decision Tree - explainable rules

Random Forest / Gradient Boosting - ensembles

About this tool

The concept

How it works

Winning strategy