Logistic Regression for Classification

While the name "Logistic Regression" includes "Regression," it's actually one of the most fundamental and widely used algorithms for classification tasks. Logistic Regression adapts a linear model approach to predict probabilities for categorical outcomes. Instead of predicting a continuous value, it estimates the probability that an instance belongs to a particular class.

From Linear Regression to Probabilities: The Sigmoid Function

Recall that a standard linear regression model predicts an output $y$ using a linear combination of input features $x$:

$\hat{y} = w^T x + b$

The output $\hat{y}$ can range from $-\infty$ to $+\infty$. This is suitable for regression, but for classification, we need an output that represents a probability, constrained between 0 and 1. How can we map the potentially unbounded output of the linear equation to this range?

This is where the sigmoid function, also known as the logistic function, comes into play. It takes any real-valued number $z$ and squashes it into the range (0, 1). The sigmoid function is defined as:

$g(z) = \frac{1}{1 + e^{-z}}$

Here, $z$ is typically the output of the linear part of the model, so $z = w^T x + b$. The function $g(z)$ gives us the estimated probability $P(y=1 | x; w, b)$ that the output $y$ is 1, given the input features $x$ and the learned parameters $w$ (weights) and $b$ (bias).

Let's visualize the sigmoid function:

The sigmoid function maps any input value $z$ to an output between 0 and 1. As $z$ becomes very large positive, $g(z)$ approaches 1. As $z$ becomes very large negative, $g(z)$ approaches 0. When $z=0$, $g(z)=0.5$.

Making Predictions: The Decision Boundary

The output of the sigmoid function gives us a probability. To make a concrete class prediction (e.g., 0 or 1 for binary classification), we need a decision threshold. A common choice is 0.5.

• If $g(z) \ge 0.5$, predict class 1.
• If $g(z) < 0.5$, predict class 0.

Looking at the sigmoid function plot, we see that $g(z) \ge 0.5$ happens when $z \ge 0$. Since $z = w^T x + b$, our decision rule becomes:

• Predict 1 if $w^T x + b \ge 0$
• Predict 0 if $w^T x + b < 0$

The equation $w^T x + b = 0$ defines the decision boundary. For Logistic Regression with this linear form of $z$, the decision boundary is a line (in 2D), a plane (in 3D), or a hyperplane (in higher dimensions) that separates the two classes.

An example scatter plot showing two classes of data points and a linear decision boundary learned by Logistic Regression. Points on one side are classified as Class 0, and points on the other side as Class 1.

How Logistic Regression Learns: The Cost Function

Just like linear regression needs a cost function to measure how well the line fits the data, logistic regression needs one too. However, using the Mean Squared Error (MSE) from linear regression with the sigmoid function results in a non-convex cost function with many local minima. This makes it difficult for optimization algorithms like gradient descent to find the global minimum reliably.

Instead, Logistic Regression uses a cost function called Log Loss (also known as Binary Cross-Entropy). For a single training example ($x^{(i)}, y^{(i)}$), where $y^{(i)}$ is the true label (0 or 1) and $h(x^{(i)}) = g(w^T x^{(i)} + b)$ is the predicted probability for class 1, the cost is:

$\text{Cost}(h(x^{(i)}), y^{(i)}) = -[y^{(i)} \log(h(x^{(i)})) + (1 - y^{(i)}) \log(1 - h(x^{(i)}))]$

The total cost function $J(w, b)$ over all $m$ training examples is the average of these individual costs:

$J(w, b) = \frac{1}{m} \sum_{i=1}^{m} \text{Cost}(h(x^{(i)}), y^{(i)})$ $J(w, b) = -\frac{1}{m} \sum_{i=1}^{m} [y^{(i)} \log(h(x^{(i)})) + (1 - y^{(i)}) \log(1 - h(x^{(i)}))]$

Let's analyze this cost:

• If the true label $y^{(i)}=1$, the cost is $-\log(h(x^{(i)}))$. This cost approaches 0 as the predicted probability $h(x^{(i)})$ approaches 1, and approaches $\infty$ as $h(x^{(i)})$ approaches 0. The model is penalized heavily for predicting a low probability when the actual class is 1.
• If the true label $y^{(i)}=0$, the cost is $-\log(1 - h(x^{(i)}))$. This cost approaches 0 as $h(x^{(i)})$ approaches 0 (meaning $1-h(x^{(i)})$ approaches 1), and approaches $\infty$ as $h(x^{(i)})$ approaches 1. The model is penalized heavily for predicting a high probability when the actual class is 0.

This Log Loss function is convex, making it suitable for optimization algorithms like gradient descent to find the optimal parameters $w$ and $b$.

Logistic Regression in Scikit-learn

Scikit-learn provides a straightforward implementation of Logistic Regression in the sklearn.linear_model.LogisticRegression class.

Here's a basic example of how to use it:

# Import necessary libraries
import numpy as np
from sklearn.model_selection import train_test_split
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import accuracy_score
from sklearn.datasets import make_classification # For generating sample data

# Generate some synthetic classification data
X, y = make_classification(n_samples=200, n_features=2, n_informative=2, n_redundant=0,
                           n_clusters_per_class=1, random_state=42, flip_y=0.1)

# Split data into training and testing sets
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3, random_state=42)

# 1. Instantiate the LogisticRegression model
#    Common parameters:
#    - C: Inverse of regularization strength (smaller C means stronger regularization). Default is 1.0.
#    - penalty: Type of regularization ('l1', 'l2', 'elasticnet', 'none'). Default is 'l2'.
#    - solver: Algorithm to use for optimization. Default often 'lbfgs'.
#             Different solvers support different penalties.
log_reg = LogisticRegression(solver='liblinear', random_state=42) # liblinear is good for smaller datasets

# 2. Train (fit) the model on the training data
log_reg.fit(X_train, y_train)

# 3. Make predictions on the test data
y_pred = log_reg.predict(X_test)

# 4. Predict probabilities on the test data
#    Returns an array of shape (n_samples, n_classes)
#    Each row sums to 1. Columns correspond to classes sorted numerically.
y_pred_proba = log_reg.predict_proba(X_test)

# Evaluate the model (we'll cover metrics in detail later)
accuracy = accuracy_score(y_test, y_pred)

print(f"Model Coefficients (w): {log_reg.coef_}")
print(f"Model Intercept (b): {log_reg.intercept_}")
print(f"Sample Predicted Probabilities (first 5):\n{y_pred_proba[:5]}")
print(f"Sample Predictions (first 5): {y_pred[:5]}")
print(f"Test Accuracy: {accuracy:.4f}")

Important points about the Scikit-learn implementation:

• Regularization: By default, LogisticRegression applies L2 regularization (controlled by the C parameter) to prevent overfitting, especially when dealing with many features. Lower values of C increase the regularization strength.
• Solvers: Different optimization algorithms (solver) are available ('liblinear', 'lbfgs', 'sag', 'saga', 'newton-cg'). Some solvers are faster for large datasets, while others support different regularization types.
• predict_proba: This method is useful when you need the actual probability estimates, not just the final class prediction. It returns an array where each row corresponds to a sample, and each column corresponds to a class, containing the probability of that sample belonging to that class.

Handling Multi-class Classification

What if you have more than two classes (e.g., {Cat, Dog, Bird})? Logistic Regression can be extended to handle multi-class problems using two main strategies:

1. One-vs-Rest (OvR): Also known as One-vs-All (OvA). This is the default strategy used by Scikit-learn's LogisticRegression. In this approach, for a problem with $K$ classes, $K$ separate binary logistic regression classifiers are trained. The first classifier predicts Class 1 vs. {All Other Classes}, the second predicts Class 2 vs. {All Other Classes}, and so on. When predicting for a new instance, all $K$ classifiers provide a probability, and the class corresponding to the classifier with the highest probability is chosen.
2. Multinomial Logistic Regression (Softmax Regression): Instead of training multiple binary classifiers, this approach generalizes logistic regression directly to handle multiple classes. It uses the softmax function (a generalization of the sigmoid) to output a probability distribution across all $K$ classes. You can enable this in Scikit-learn by setting multi_class='multinomial' along with a compatible solver like 'lbfgs' or 'newton-cg'.

For many practical purposes, the default OvR strategy often works well.

Strengths and Weaknesses

Strengths:

• Interpretability: Similar to linear regression, the coefficients ($w$) can provide insights into the importance and direction of influence of each feature on the probability outcome (log-odds).
• Efficiency: It's computationally inexpensive to train and predict, making it suitable for large datasets.
• Probabilistic Output: Provides well-calibrated probabilities, which can be useful for decision-making in classification.
• Baseline Model: Often serves as a strong baseline model to compare against more complex algorithms.

Weaknesses:

• Linearity Assumption: Assumes a linear relationship between the features and the log-odds of the outcome. It might not capture complex, non-linear patterns effectively without feature engineering (e.g., creating polynomial features).
• Sensitivity to Outliers: Like linear regression, it can be sensitive to outliers in the feature space.
• Performance on Non-linear Problems: May underperform compared to more complex models (like SVMs with kernels, or tree-based methods) when the decision boundary is highly non-linear.

Logistic Regression is a powerful yet simple classification algorithm that forms a foundation of many machine learning applications. Understanding its mechanics, including the sigmoid function, decision boundary, and cost function, provides a solid foundation before moving on to other classification techniques like K-Nearest Neighbors and Support Vector Machines, which we will discuss next.