Unit 6: Machine Learning

I. Orientation — Foundations of Machine Learning

Machine learning (ML) is a branch of artificial intelligence in which computational systems improve their performance on a task by learning patterns from data rather than relying only on explicitly programmed rules. Its modern foundations include statistical inference, optimization, computer science, and pattern recognition.

Defining properties:

• Data-driven operation: An ML system estimates relationships from examples such as labelled emails, historical sales records, images, or sensor readings.
• Generalization: The objective is not merely to memorize training data but to perform well on previously unseen examples drawn from a similar population.
• Model: A model is a parameterized representation of a learned relationship; for example, linear regression uses parameters (w) and (b) in (\hat{y}=wx+b).
• Training: Learning adjusts model parameters to reduce a measurable error or improve a reward.
• Inference: After training, the model processes new input (x) to produce a prediction (\hat{y}), class, score, action, or structured output.
• Evaluation: Performance is measured with criteria such as accuracy, mean squared error, precision, recall, or cumulative reward.
• Assumption: Training and future data must have sufficiently similar statistical properties; major distribution changes can make learned patterns unreliable.

A. Introduction

Machine learning converts experience represented by data into a model that can make predictions, discover structure, or select actions.

• Classical programming contrast:
  1. Rule-based programming: A programmer supplies data and explicit rules to produce answers.
  2. Machine learning: A learning algorithm receives data and desired answers, rewards, or structural criteria and produces a model containing learned rules.
• Core mapping: Many learning tasks estimate a function (f) connecting input features to an output.

ŷ = f(x; θ)

• (x): input or feature vector, such as a house’s area and age.
• (\theta): parameters learned from data, such as regression coefficients.
• (f): model or prediction function.
• (\hat{y}): predicted output, such as house price.
  • Typical workflow: Data collection is followed by cleaning, feature preparation, model training, validation, testing, deployment, and monitoring.
  • Generalization requirement: A model with 99% training accuracy but 70% test accuracy has probably learned training-specific noise rather than a dependable relationship.
  • Role in data science: ML turns analyzed data into operational outputs, including fraud alerts, demand forecasts, medical-image classifications, and product recommendations.

II. Types of Machine Learning Techniques — Sources of Learning Signals

Machine learning techniques are primarily distinguished by the information available during training: correct labels, no labels, a mixture of both, or feedback obtained through interaction.

A. Types of machine learning techniques

The principal learning types differ in their training signals, objectives, and suitable applications.

1. Supervised learning: The algorithm learns from labelled pairs ((x_i,y_i)), where (x_i) is the (i)-th input and (y_i) is its known target.
   • Classification: Predicts a discrete label; an email classifier may output “spam” or “not spam.” Common algorithms include logistic regression, decision trees, support-vector machines, and neural networks.
   • Regression: Predicts a continuous value; a model may estimate a temperature of (23.7^\circ\text{C}) or a price of ₹500,000.
   • Objective: Parameters commonly minimize average loss:

θ* = arg minθ (1/n) Σᵢ L(yᵢ, f(xᵢ; θ))

• (n): number of training examples.
• (L): loss measuring prediction error.
• (\theta^*): parameter values producing the smallest training objective.

2. Unsupervised learning: The algorithm receives inputs (x_i) without known targets and searches for useful structure.

   • Clustering: Groups similar observations; (k)-means assigns customers to (k) groups by minimizing distances from cluster centers.
   • Dimensionality reduction: Methods such as principal component analysis represent many correlated variables with fewer components.
   • Association discovery: Market-basket analysis can identify products frequently purchased together.
   • Limitation: Without labels, clusters are not automatically meaningful; a three-cluster solution requires interpretation using domain variables.

3. Semi-supervised learning: A small labelled dataset is combined with a larger unlabelled dataset.

   • Concrete setting: A system may use 1,000 labelled images and 50,000 unlabelled images when expert annotation is expensive.
   • Mechanism: Techniques include pseudo-labelling, consistency regularization, and graph-based propagation.
   • Risk: Incorrect pseudo-labels can be reinforced, so confidence thresholds and validation data are important.

4. Reinforcement learning: An agent learns actions through interaction with an environment and receives numerical rewards.

   • Elements: At time (t), the agent observes state (s_t), chooses action (at), receives reward (r{t+1}), and reaches state (s_{t+1}).
   • Objective: The policy seeks to maximize expected cumulative discounted reward:

Gₜ = rₜ₊₁ + γrₜ₊₂ + γ²rₜ₊₃ + ...

• (G_t): return from time (t).
• (\gamma): discount factor between 0 and 1.
• Applications: Robot control, game playing, resource allocation, and sequential recommendation.
• Challenge: Exploration discovers new actions, whereas exploitation selects actions already believed to be effective.

III. Learning Problems and Systems — From Data to Generalization

A learning problem formally specifies what must be learned, while a learning system includes the data, representation, algorithm, computing process, and evaluation mechanism needed to learn it.

A. Learning problems and system

A well-defined learning problem connects a task, a source of experience, and an explicit performance criterion.

• Three-part formulation: A program learns from experience (E) with respect to task (T) and performance measure (P) if its measured performance at (T) improves with (E).
  • Task (T): Classify incoming messages as spam or legitimate.
  • Experience (E): A dataset of previously classified messages.
  • Performance (P): Accuracy, precision, recall, or another metric on unseen messages.
• Common learning problems:
  • Classification: Map (x) to one of (K) classes, such as digits 0–9.
  • Regression: Estimate a numerical target, such as tomorrow’s electricity demand in megawatts.
  • Ranking: Order documents according to relevance to a search query.
  • Clustering: Partition unlabelled observations according to similarity.
  • Anomaly detection: Identify unusual cases, such as a transaction far from normal account behavior.
• System components:
  • Data pipeline: Collects, validates, transforms, and divides data.
  • Feature representation: Converts raw objects into usable variables; a document may become a vector of word frequencies.
  • Learning algorithm: Searches a hypothesis space for suitable parameters.
  • Model repository: Stores trained parameters, versions, and metadata.
  • Serving layer: Exposes predictions through a batch process or application programming interface.
  • Monitoring layer: Tracks latency, errors, input drift, and post-deployment performance.
• Generalization error: The expected loss on unseen population data is more important than training loss, but it must be estimated with held-out data.
• Data partitioning: A typical split might allocate 70% for training, 15% for validation, and 15% for final testing; the test set remains untouched until model selection is complete.
• Central failure modes:
  1. Underfitting: The model is too simple and performs poorly on both training and test data.
  2. Overfitting: The model performs well on training data but poorly on new data because it captures noise.
• Data leakage: Using future or test information during training creates misleading results; including a “payment recovered” field in a default-risk model would reveal an outcome unavailable at prediction time.

IV. Designing a Learning System — An End-to-End Engineering Process

Designing a learning system requires aligning the model with the real decision, available data, operating constraints, and consequences of errors.

A. Designing a learning system

Effective design proceeds from problem definition to deployment and continuous monitoring rather than beginning with algorithm selection.

• Define the objective: State the input, output, prediction horizon, users, and decision. “Predict whether an invoice will be unpaid 30 days after its due date” is more precise than “predict bad customers.”
• Choose a performance measure: Accuracy may be unsuitable for rare events; if only 1% of transactions are fraudulent, predicting “not fraud” always gives 99% accuracy but detects no fraud.
• Collect representative data: Sampling should cover relevant users, seasons, devices, and operating conditions; missing an entire region can create systematic deployment errors.
• Prepare data: Handle duplicates, missing values, inconsistent units, categorical variables, and extreme values using transformations learned from training data only.
• Select features and representation: Domain variables may be combined with learned representations; transaction amount, hour, location distance, and device history can support fraud detection.
• Choose a model family: Consider predictive quality, interpretability, training cost, inference speed, memory, and maintainability. A small decision tree may be preferred when decisions require clear explanations.
• Train and tune: Fit parameters on training data and choose hyperparameters—such as tree depth or regularization strength—using validation data or cross-validation.
• Evaluate realistically: Use a time-based split for forecasting and grouped splits when records from one person must not appear in both training and test sets.
• Deploy safely: Package preprocessing and the model together, version them, log predictions, and provide fallback behavior for missing or malformed inputs.
• Monitor and retrain: Compare current feature distributions with training distributions and measure delayed outcomes. A changing relationship between inputs and targets indicates concept drift.
• Account for responsible use: Examine subgroup error rates, privacy, security, explainability, and human oversight; aggregate accuracy can conceal high false-negative rates for a smaller group.

V. Concept of Learning Task — Defining What Improvement Means

A learning task is a precise computational objective in which experience is used to improve measurable performance under stated conditions.

A. Concept of learning task

The concept of a learning task separates the desired capability from the algorithm chosen to implement it.

• Task specification: Define the input space (X), output space (Y), target relationship, and loss function. For binary classification, (Y={0,1}).
• Example representation: In credit-risk prediction, (x\in X) may contain income, debt ratio, and repayment history, while (y=1) denotes default.
• Hypothesis space: The learner searches a set (H) of possible functions (h:X\rightarrow Y); restricting (H) controls what patterns can be represented.
• Loss and risk: For binary classification, zero-one loss records whether a prediction is wrong:

L(y, ŷ) = 0 if y = ŷ; otherwise 1

• (y): actual class.
• (\hat{y}): predicted class.
• (L): error assigned to one prediction.
  • Inductive learning: The system infers a general rule from finite examples; this requires an inductive bias such as preference for simpler functions or nearby cases.
  • Performance distinction: Training performance measures fit to observed examples, whereas test performance estimates success on unseen examples.
  • Success condition: Learning has occurred only when additional experience improves the chosen performance measure on the specified task, not merely when a complex model has been fitted.