Machine Learning Basics: Models, Algorithms & Evaluation

What is Machine Learning?

Machine Learning (ML) is a subset of AI where systems learn patterns from data without explicit programming. It relies on algorithms that improve performance with experience (more data). ML is categorized into three main types:

1. Supervised Learning

• Definition: The model learns from labeled data (input-output pairs).

• Examples: Spam detection (email labeled as spam/not spam), house price prediction (features: size, location; output: price).

• Types:

  • Classification: Predicts discrete labels (e.g., "cat" or "dog").
  • Regression: Predicts continuous values (e.g., temperature, stock price).

  flowchart LR
    A["Supervised Learning"] --> B["Classification\n(e.g., Email Spam)"]
    A --> C["Regression\n(e.g., House Price)"]

2. Unsupervised Learning

• Definition: The model finds hidden patterns in unlabeled data.

• Examples: Customer segmentation (grouping users by behavior), anomaly detection (fraud in transactions).

• Types:

  • Clustering: Groups similar data (e.g., K-means).
  • Dimensionality Reduction: Simplifies data (e.g., PCA).

  flowchart LR
    A["Unsupervised Learning"] --> B["Clustering\n(e.g., Customer Groups)"]
    A --> C["Dimensionality Reduction\n(e.g., PCA)"]

3. Reinforcement Learning

• Definition: The model learns by interacting with an environment (rewards/punishments).
• Examples: Self-driving cars (reward: safe navigation), Pathao’s dynamic pricing (reward: high demand).

Key ML Algorithms

Model Evaluation Metrics

1. For Classification

Worked Example (Confusion Matrix for Spam Detection) Suppose:

• True Positives (TP): 80 (correctly identified spam)
• False Positives (FP): 10 (ham marked as spam)
• False Negatives (FN): 20 (spam missed)
• True Negatives (TN): 900 (correctly identified ham)

Calculate:

• Accuracy = $\frac{80+900}{80+900+10+20}=96.2\%$
• Precision = $\frac{80}{80+10}=88.9\%$
• Recall = $\frac{80}{80+20}=80\%$

2. For Regression

Worked Example (Predicting Daraz Delivery Time) Suppose actual vs. predicted delivery times (hours):

MSE = $\frac{0.01+0.01+0.25}{3}=0.09$

Bias-Variance Tradeoff

• Bias: Error due to overly simplistic assumptions (underfitting). Example: Linear regression failing to capture nonlinear trends in stock prices.
• Variance: Error due to excessive sensitivity to small data fluctuations (overfitting). Example: A decision tree with 20 levels memorizing training data but failing on new data.

graph LR
  A["High Bias\n(Underfitting)"] --> B["Simple Model\n(High Error on Train & Test)"]
  C["Low Bias\n(Good Fit)"] --> D["Balanced Model\n(Low Error on Both)"]
  E["High Variance\n(Overfitting)"] --> F["Complex Model\n(Low Train Error, High Test Error)"]

Solution: Use techniques like:

• Cross-validation (split data into training/validation sets).
• Regularization (penalize complexity, e.g., L1/L2 norms).
• Ensemble methods (combine multiple models, e.g., Random Forest).

Real-World Applications in Nepal

1. eSewa (Fraud Detection)

• Idea Used: Supervised Learning (Classification)
• How: eSewa uses ML to flag suspicious transactions (e.g., unusual payment patterns) by training on labeled fraud/non-fraud data.
• Algorithm: Random Forest or Logistic Regression.
• Metric: High recall (catch most frauds) is prioritized over precision (some false alarms are acceptable).

2. Pathao (Dynamic Pricing)

• Idea Used: Reinforcement Learning
• How: Pathao adjusts ride prices in real-time based on demand/supply (reward: maximizing driver earnings and passenger satisfaction).
• Algorithm: Q-Learning or Deep Q-Networks (DQN).
• Visual:

  flowchart LR
    A["High Demand\n(Low Supply)"] --> B["Increase Price\n(Reward: More Drivers)"]
    C["Low Demand\n(High Supply)"] --> D["Decrease Price\n(Reward: More Riders)"]

3. NTC (Network Traffic Prediction)

• Idea Used: Time-Series Forecasting (Regression)
• How: NTC predicts internet traffic spikes (e.g., during exams) to optimize bandwidth allocation.
• Algorithm: ARIMA or LSTM (for sequential data).
• Worked Example: Suppose NTC’s historical daily traffic (in GB):

  Day	Traffic
  1	1000
  2	1200
  3	1100

  Linear Regression Model: $\hat{y}=1050+50x$
  • Predict Day 4: $\hat{y}=1050+50(3)=1200$ GB.

Exam Tip

1. Understand the difference between supervised/unsupervised learning—exams often ask for examples.
2. Practice confusion matrices: Given TP/FP/FN/TN, calculate accuracy, precision, and recall.
3. Bias-variance tradeoff: Know how to diagnose underfitting/overfitting (e.g., high training error = high bias; low test error but high training error = high variance).
4. Real-world mapping: Relate algorithms to Nepalese apps (e.g., "Which ML technique does Pathao use for pricing?").
5. Visuals matter: Draw confusion matrices or bias-variance curves in exams to explain answers.

A labeled flowchart showing data collection → preprocessing → model training → evaluation → deployment. (Image: Generated and edited with Genspark (Nano Banana 2); prompt d, Public domain, via Wikimedia Commons) A simple 3-layer neural network (input → hidden → output) with weights and activation functions. (Image: BrunelloN, CC BY-SA 4.0, via Wikimedia Commons)