Regressione logistica e funzione sigmoide

Logistic regression is the go-to algorithm for binary classification. Despite its name, it is a classifier, not a regressor. It extends the linear model by squashing the output through a special function called the sigmoid, ensuring predictions are always valid probabilities between 0 and 1.

Logistic regression is one of the most widely deployed ML models in production. It is fast to train, interpretable, and provides calibrated probability estimates. It is often the first classifier you should try on a new binary classification problem before reaching for more complex methods.

The sigmoid function (also called the logistic function) maps any real number to the range (0, 1):

σ(z) = 1 / (1 + e^(-z))

Key properties: σ(0) = 0.5, σ(+∞) → 1, σ(-∞) → 0. The S-shaped curve smoothly transitions from 0 to 1. When z is a large positive number (strong signal for class 1), the output approaches 1. When z is a large negative number (strong signal for class 0), the output approaches 0.

Logistic regression first computes the same linear combination as linear regression (z = w₁x₁ + w₂x₂ + ... + b), then passes it through the sigmoid to get a probability:

P(y=1|x) = σ(wᵀx + b) = 1 / (1 + e^(-(wᵀx+b)))

The model outputs a probability that the input belongs to class 1. If this probability exceeds 0.5, the model predicts class 1; otherwise class 0. The weights w are learned during training to maximise the likelihood of the observed labels — a process called maximum likelihood estimation.

Logistic regression does not use MSE as its cost function. Instead, it uses Binary Cross-Entropy (log loss):

Loss = -(y × log(ŷ) + (1-y) × log(1-ŷ))

This penalises confident wrong predictions extremely harshly. If the true label is 1 and the model predicts probability 0.001, the log loss is -log(0.001) ≈ 6.9 — enormous. If the model predicts 0.99, the loss is -log(0.99) ≈ 0.01 — tiny. This asymmetric penalisation is exactly right for probability calibration.

Scikit-learn's LogisticRegression uses a gradient-based optimiser (LBFGS by default) to minimise log loss and find the optimal weights. The API is identical to LinearRegression: instantiate, fit, predict.

Important parameters include C (inverse of regularisation strength — smaller C means stronger regularisation) and max_iter (maximum number of optimisation steps). If you see a convergence warning, increase max_iter or scale your features first.

Logistic regression is one of the few classifiers that produces well-calibrated probability estimates. A probability of 0.85 for class 1 means roughly 85% of predictions with that confidence level should actually be class 1. This calibration is valuable for risk management applications (credit scoring, medical diagnosis) where the probability itself matters, not just the binary label.

Use predict_proba() to retrieve the full probability vector. The two columns represent P(class 0) and P(class 1). They always sum to 1.0.

Logistic regression coefficients are not as directly interpretable as linear regression coefficients, but they still carry meaningful information. The coefficient w for a feature represents the change in the log-odds of the positive class for a one-unit increase in that feature:

log(P(y=1)/P(y=0)) = wᵀx + b

The odds ratio for a feature is exp(w). If exp(w) = 2, that feature doubles the odds of the positive class. Large positive coefficients indicate features that strongly predict class 1; large negative coefficients strongly predict class 0.

Logistic regression in scikit-learn applies L2 regularisation by default, controlled by the parameter C. Unlike most regularisation parameters, C is the inverse of regularisation strength: a smaller C means stronger regularisation (more shrinkage of weights), and a larger C means weaker regularisation.

Use penalty='l1' and solver 'liblinear' for L1 regularisation, which performs automatic feature selection by zeroing out irrelevant feature weights. L1 logistic regression is particularly useful when you have many features and suspect most are irrelevant.

A single accuracy number hides important information. Scikit-learn's classification_report() prints precision, recall, and F1-score for each class, plus macro and weighted averages. This is the standard way to report classification performance in research and industry.

Reading the report: the 'support' column shows how many examples of each class are in the test set. If supports are very unequal (class imbalance), look at per-class recall rather than overall accuracy to understand where the model fails.

A clear comparison of when to use each:

• Target type: Linear regression → continuous number; Logistic regression → probability / binary label.
• Output range: Linear → (-∞, +∞); Logistic → (0, 1).
• Loss function: Linear → MSE; Logistic → Binary Cross-Entropy (log loss).
• Evaluation: Linear → RMSE, R²; Logistic → Accuracy, F1, AUC-ROC.
• Shared property: Both are linear models — the decision boundary is a hyperplane, and both benefit from feature scaling and regularisation.

For problems with more than two classes, logistic regression generalises to Softmax regression (also called multinomial logistic regression). Instead of one sigmoid, it computes one linear score per class and passes them all through the softmax function, producing a probability distribution over all classes.

In scikit-learn, set multi_class='multinomial' and use solver 'lbfgs' or 'saga'. The model outputs a probability for each class, and the predicted class is the one with the highest probability.

In this lesson you learned: the sigmoid function squashes any linear score to a valid probability between 0 and 1, logistic regression uses binary cross-entropy loss which harshly penalises confident wrong predictions, and coefficients represent changes in log-odds with odds ratio exp(w) giving a more intuitive scale-invariant interpretation. Next up we build a confusion matrix from predictions and ground truth, providing a detailed breakdown of where the classifier succeeds and fails.