Core supervised and unsupervised learning, model evaluation, and the basics of neural networks, aimed at technical interview prep and equivalent coursework.
Every skill the diagnostic can check, grouped the way the map is. Open any topic to start a lesson on it.
Foundations
Types of ML problems: The basic split between supervised (labeled examples), unsupervised (no labels), and reinforcement learning.
Cost functions and loss: The single number a model tries to minimize during training, chosen to reflect how wrong its predictions are.
Gradient descent: Repeatedly nudging a model's parameters downhill on the cost function to reduce error.
Supervised learning
Linear regression: Fitting a straight line (or hyperplane) through data to predict a continuous number.
Logistic regression: A linear model adapted for classification by squashing its output into a probability.
Decision trees: A model that predicts by asking a sequence of yes/no questions about the input, like a flowchart.
Ensemble methods: Combining many weak models (usually trees) into one stronger one, either by averaging (bagging) or correcting each other's mistakes in sequence (boosting).
Support vector machines: A classifier that finds the widest possible margin between classes, and can bend that margin into curves using kernels.
K-nearest neighbors: Classifying a new point by looking at whatever its closest labeled neighbors are.
Naive Bayes: A probabilistic classifier that (naively) assumes every feature is independent of the others given the class.
Model evaluation
Bias-variance tradeoff: A model too simple has high bias (underfits); one too flexible has high variance (overfits), and you can't minimize both at once.
Overfitting and underfitting: A model that memorizes training data instead of learning the pattern (overfit), or one too weak to learn it at all (underfit).
Train/test/validation split: Holding out data the model never trains on, so you can honestly check how it does on unseen examples.
Cross-validation: Splitting the data multiple different ways and averaging performance, so one lucky/unlucky split doesn't mislead you.
Regularization (L1/L2): Penalizing large model weights during training so it can't overfit by memorizing noise.
Classification evaluation metrics: Accuracy alone can lie; precision, recall, F1, and ROC-AUC each capture a different way a classifier can be right or wrong.
Regression evaluation metrics: Ways to score how far off a model's continuous predictions are, from raw error size to how much variance it explains.
Hyperparameter tuning: Systematically searching for the settings (learning rate, regularization strength, tree depth) that aren't learned from data but chosen before training.
Class imbalance handling: Techniques for when one class vastly outnumbers another, since accuracy alone becomes meaningless (predicting the majority class always "wins").
Feature engineering
Feature engineering and scaling: Reshaping raw data (scaling numbers, encoding categories) into a form a model can actually learn from well.
Unsupervised learning
Dimensionality reduction (PCA): Compressing many correlated features into fewer new ones that still capture most of the original variation.
K-means clustering: Grouping unlabeled points into k clusters by repeatedly assigning them to the nearest center and recomputing centers.
Hierarchical clustering: Building a tree of nested clusters by repeatedly merging (or splitting) groups, instead of picking a fixed number upfront.
Neural networks
Neural network basics: Stacking layers of logistic-regression-like units, connected by non-linear activation functions, to learn more complex patterns.
Backpropagation: The algorithm that computes how much each weight in a neural network contributed to the error, so gradient descent knows which way to adjust it.