📊 Logistic Regression for Classification

In classification problems, the output variable $y$ takes discrete values

Classification Types

1. Binary Classification

Two classes:

$$y \in \{0,1\}$$

We usually call:

• $0$ → Negative class: 0 represents absence.
• $1$ → Positive class: 1 represent presence of something (e.g., disease)

2. Multi-class Classification

More than two classes:

$$y \in \{0,1,2,3,...\}$$

---

The Sigmoid Function $\sigma(x)$

Sigmoid function (also called logistic function) maps any real-valued number into the (0, 1) interval.

• It is commonly used in logistic regression to model probabilities.

The sigmoid function is defined as:

$$\sigma(x) = \frac{1}{1 + e^{-x}} = \frac{e^x}{e^x + 1} = 1 - \frac{1}{1 + e^x} = \frac{1}{2} \left(1 + \tanh\left(\frac{x}{2}\right)\right) = 1 - \sigma(-x)$$

Where:

• $z$ is the input to the function (can be any real number)
• $e$ is the base of the natural logarithm (approximately 2.71828)

Output:

$\sigma(z)$ is always between 0 and 1, making it suitable for modeling probabilities.

• When $z$ is large and positive, $\sigma(z) \approx 1$.
  • $z \to +\infty$, $\sigma(z) \to 1$
• When $z$ is large and negative, $\sigma(z) \approx 0$.
  • $z \to -\infty$, $\sigma(z) \to 0$
• When $z = 0$, $\sigma(z) = 0.5$.

---

💡 Logistic Regression Hypothesis $h_\theta(x)$

Logistic regression ensures:

$$0 \le h_\theta(x) \le 1$$

Where

• Input: Any real number : $(-\infty, +\infty)$
• Output: Always between: $(0,1)$

Instead of: $h_\theta(x) = \theta^T x$

We apply a transformation that squashes outputs into the probability range $[0,1]$.

$$h_\theta(x) = g(\theta^T x)$$

So the output becomes a probability: $h_\theta(x) = P(y=1 \mid x)$

This can be simplified to:

$$h_\theta(x) = g(z)$$

Where

$$z = \theta^T x$$

and $g(z)$ as the sigmoid function:

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

Final Hypothesis

$$h_\theta(x) = \frac{1}{1 + e^{-\theta^T x}}$$

This ensures:

$$0 \le h_\theta(x) \le 1$$ $$h_\theta(x) = P(y = 1 \mid x; \theta)$$

So:

• If $h_\theta(x) = 0.7$ → There is a 70% probability that $y = 1$

Since probabilities must sum to 1:

$$P(y=0 \mid x; \theta) = 1 - P(y=1 \mid x; \theta)$$

---

Decision Boundary

The decision boundary is the line that separates the area where y = 0 and where y = 1.

• It is created by our hypothesis function / model.

Decision Boundary is a Property of the Model

The decision boundary depends only on:

• The hypothesis form
• The parameters $\theta$

It does not depend on the training data once $\theta$ is fixed.

The training set is used only to learn $\theta$.

Suppose we have a Classification Rule

In order to get our discrete 0 or 1 classification, we can translate the output of the hypothesis function as follows:

• ➕ $h_\theta(x) \ge 0.5$ we Predict $y = 1$
• ➖ $h_\theta(x) < 0.5$ we Predict $y = 0$

When Is $h_\theta(x) \ge 0.5$?

Since:

$$g(z) \ge 0.5 \quad \text{when} \quad z \ge 0$$

and

$$h_\theta(x) = g(\theta^T x),$$

we predict:

$$y = 1 \quad \text{when} \quad \theta^T x \ge 0$$

and

$$y = 0 \quad \text{when} \quad \theta^T x < 0$$

Linear Decision Boundary

Suppose:

$$h_\theta(x) = g(\theta_0 + \theta_1 x_1 + \theta_2 x_2)$$

Let:

$$\theta_0 = -3, \quad \theta_1 = 1, \quad \theta_2 = 1$$

Then:

$$\theta^T x = -3 + x_1 + x_2$$

We predict $y = 1$ when:

$$-3 + x_1 + x_2 \ge 0$$

Rewriting:

$$x_1 + x_2 \ge 3$$

Decision Boundary

The decision boundary occurs when:

$$x_1 + x_2 = 3$$

This is a straight line.

It separates the plane into:

• Region where $y = 1$
• Region where $y = 0$

The decision boundary corresponds to:

$$h_\theta(x) = 0.5$$

---

Nonlinear Decision Boundaries

We can add polynomial features.

Example:

$$h_\theta(x) = g(\theta_0 + \theta_1 x_1+ \theta_2 x_2+ \theta_3 x_1^2+ \theta_4 x_2^2)$$

Suppose:

$$\theta_0 = -1, \quad \theta_1 = 0, \quad \theta_2 = 0, \quad \theta_3 = 1, \quad \theta_4 = 1$$

Then:

$$\theta^T x = -1 + x_1^2 + x_2^2$$

We predict $y = 1$ when:

$$-1 + x_1^2 + x_2^2 \ge 0$$

Rewriting:

$$x_1^2 + x_2^2 \ge 1$$

Decision Boundary

The boundary is:

$$x_1^2 + x_2^2 = 1$$

This is a circle of radius 1.

So logistic regression can produce nonlinear boundaries using polynomial features.

---

More Complex Boundaries

By adding higher-order terms such as:

• $x_1^3$
• $x_1 x_2$
• $x_1^2 x_2$
• etc.

Logistic regression can represent:

• Ellipses
• Complex curves
• Highly nonlinear shapes

---

💰 Cost Function / Optimal Objective

The overall cost is:

$$J(\theta) = \frac{1}{m} \sum_{i=1}^{m} \text{Cost}\big(h_\theta(x^{(i)}), y^{(i)}\big)$$

where:

$$h_\theta(x) = \frac{1}{1 + e^{-\theta^T x}}$$

Why Not Use Squared Error Cost?

In linear regression, we use:

$$J(\theta) = \frac{1}{2m}\sum (h_\theta(x) - y)^2$$

If we use same squared error with sigmoid:

• The cost function becomes non-convex
• Optimization may get stuck in local minima
• Training may fail to find the best parameters

So we need a better cost function.

We define cost separately for the two classes.

The cost function is defined as:

$$\text{Cost}(h_\theta(x), y) = \begin{cases} -\log(h_\theta(x)) & \text{if } y = 1 \\ -\log(1 - h_\theta(x)) & \text{if } y = 0 \end{cases}$$

Why This Cost Function Is Better

It is convex

• No local minima
• Guarantees only one global minimum
• Optimization is reliable
• Smooth and well-behaved

Because $J(\theta)$ is convex:

• Gradient descent will converge to the global minimum
• We do not get stuck in bad local optima
• Training is stable

It penalizes wrong predictions heavily

• Cost = 0 when prediction is correct
• Cost → ∞ when prediction is very wrong
• Encourages the model to be confident and correct

1. Case 1: When $y = 1$

We want $h_\theta(x)$ to be close to 1.

Cost:

$$-\log(h_\theta(x))$$

If prediction is close to 1 → cost is small

• If $h_\theta(x) = 1$ → cost = 0

If prediction is close to 0 → cost is very large -If $h_\theta(x) \to 0$ → cost $\to \infty$

So:

• Correct confident prediction → small cost
• Wrong confident prediction → very large cost

Case 2: When $y = 0$

We want $h_\theta(x)$ to be close to 0.

Cost:

$$-\log(1 - h_\theta(x))$$

If prediction is close to 1 → cost is very large -If $h_\theta(x) \to 1$ → cost $\to \infty$

If prediction is close to 0 → cost is small

• If $h_\theta(x) = 0$ → cost = 0

Again:

• Correct prediction → small cost
• Wrong confident prediction → large penalty

---

Unified Logistic Cost Function

Simplified Cost Function (Single Formula)

We can combine the two cases into one equation:

$$\text{Cost}(h_\theta(x), y) = - y \log(h_\theta(x))- (1 - y)\log(1 - h_\theta(x))$$

• If $y = 1$:
  • The second term becomes 0
  • Cost reduces to:

$$-\log(h_\theta(x))$$

• If $y = 0$:
  • The first term becomes 0
  • Cost reduces to:

$$-\log(1 - h_\theta(x))$$

So this single formula covers both cases.

---

Full Cost Function

Full Cost Function Over Dataset

For $m$ training examples:

$$J(\theta) = -\frac{1}{m} \sum_{i=1}^{m} \left[ y^{(i)} \log(h_\theta(x^{(i)})) + (1 - y^{(i)}) \log(1 - h_\theta(x^{(i)})) \right]$$

This is called:

• Log loss
• Cross-entropy loss
• Logistic loss

Where:

$$h_\theta(x) = \frac{1}{1 + e^{-\theta^T x}}$$

---

Vectorized Cost Fucntion

Let:

• $X$ = design matrix
• $y$ = vector of labels
• $h = g(X\theta)$

Then:

$$h = g(X\theta)$$

and

$$J(\theta) = \frac{1}{m} \left(- y^T \log(h)- (1 - y)^T \log(1 - h) \right)$$

---

🧠 Key Takeaways: Cost Function

• Logistic regression uses a convex cost function
• The simplified cost formula works for both $y=0$ and $y=1$
• Gradient descent update looks the same as linear regression
• Vectorization makes implementation efficient
• Always include the $\frac{1}{m}$ factor in the gradient update

---

🎢 Gradient Descent

General gradient descent:

Repeat:

$$\theta_j := \theta_j- \alpha \frac{\partial}{\partial \theta_j} J(\theta)$$

Where:

• $\alpha$ = learning rate
• $J(\theta)$ = cost function

Logistic Regression Gradient

After computing the derivative, we get:

$$\theta_j := \theta_j - \frac{\alpha}{m} \sum_{i=1}^{m} \left( h_\theta(x^{(i)}) - y^{(i)} \right) x_j^{(i)}$$

Important Notes

• This is identical in form to linear regression gradient descent.
• We must update all $\theta_j$ simultaneously..
• The difference lies in the hypothesis function:

$$h_\theta(x) = g(\theta^T x)$$

where

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

---

Vectorized Gradient Descent

Let:

• $X$ = design matrix
• $\vec{y}$ = vector of labels
• $h = g(X\theta)$

Then the update rule becomes:

$$h = g(X\theta)$$

Then the update rule becomes:

$$\theta := \theta- \frac{\alpha}{m} X^T \left( g(X\theta) - \vec{y} \right)$$

Where:

• $\vec{y}$ is the vector of labels
• $X^T$ is the transpose of the design matrix

---

🧠 Key Takeaways: Gradient Descent

• Logistic regression uses gradient descent just like linear regression.
• The update formula is structurally the same.
• The cost function is different.
• The model is convex, so gradient descent converges to the global minimum.
• Vectorized form makes implementation efficient and clean.

---