ML Concepts

2. Supervised Learning

2.1 🎯 What is Supervised Learning?
1. In supervised learning, a model learns from labeled data. Labeled data contains: input variables, the correct output. We provide: X = Features or inputs y = Label or output The model learns a relationship between X and y. Scikit-learn organizes supervised-learning methods into areas such as linear models, support vector machines, nearest neighbors, decision trees, and ensemble methods.
2.2 Supervised-Learning Examples
1. House-Price Prediction: Features: house size, location, number of rooms, age of the property. Label: house price. Spam Detection: Features: email content, sender information, number of links, email metadata. Label: spam or not spam. 📌 In supervised learning, the model learns using examples where the correct output is already known.
2.3 Common Supervised-Learning Algorithms
1. Examples include: Linear Regression, Logistic Regression, K-Nearest Neighbors, Support Vector Machines, Decision Trees, Random Forests. Many algorithms can support more than one task. For example, nearest-neighbor, decision-tree, and support-vector methods can be used for classification or regression depending on the selected estimator.
2.4 Main Supervised-Learning Tasks
1. Supervised learning commonly includes two prediction tasks: regression, classification.

3. Regression

3.1 📈 What is Regression?
1. Regression is a supervised-learning task used to predict a numeric value. The output is usually a continuous number. Examples: house price, salary, temperature, sales amount, delivery time, electricity consumption. 📌 Regression answers questions such as "how much?" or "what numeric value?"
3.2 Regression Example
1. Suppose we have employee data:

3.2 Regression Example (continued)
1. Features: academic qualification, experience, company, position. Label: salary. The task is to predict the salary of the final employee.
3.3 Common Regression Algorithms
1. Examples include: Linear Regression, K-Nearest Neighbors Regression, Decision Tree Regression, Random Forest Regression, Support Vector Regression.

4. Classification

4.1 🏷️ What is Classification?
1. Classification is a supervised-learning task used to predict a category or class label. Examples: spam or not spam, disease or no disease, customer will leave or stay, transaction is fraudulent or legitimate, product category. 📌 Classification answers questions such as "which category?" or "which class?"
4.2 Classification Example
1. Suppose we have health data:

4.2 Classification Example (continued)
1. Features: age, blood pressure, cholesterol level. Label: diabetes status. The task is to predict whether the final patient belongs to the Yes or No class.
4.3 Binary and Multiclass Classification
1. Binary Classification: The output contains two possible classes. Examples: spam or not spam, disease or no disease, yes or no. Multiclass Classification: The output contains more than two possible classes. Examples: low, medium, or high risk; setosa, versicolor, or virginica; bronze, silver, or gold membership.
4.4 Common Classification Algorithms
1. Examples include: Logistic Regression, K-Nearest Neighbors, Support Vector Machines, Decision Trees, Random Forests, Naive Bayes. 📌 Despite its name, Logistic Regression is mainly used for classification.

5. Regression vs. Classification

5.1 Quick Comparison
1. A quick comparison of the two tasks:

5.2 How to Identify the Task
1. Ask: What am I trying to predict? If the answer is a numeric value: Regression If the answer is a category: Classification 📌 The target variable usually determines whether the problem is regression or classification.

6. Unsupervised Learning

6.1 🔍 What is Unsupervised Learning?
1. In unsupervised learning, a model works with unlabeled data. We provide features: X = Input variables We do not provide a known label: No y value is provided The model attempts to discover structure or patterns in the data.
6.2 Unsupervised-Learning Goals
1. Unsupervised learning can be used to: group similar observations, reduce the number of variables, discover hidden structures, find associations, support exploratory analysis.
6.3 Example: Customer Segmentation
1. Suppose we have: customer age, number of purchases, average spending, visit frequency. We do not have predefined customer segments. A clustering algorithm may identify groups such as: frequent high-value customers, occasional customers, low-spending customers. 📌 The algorithm discovers groups from similarities in the available features.
6.4 Common Unsupervised-Learning Algorithms
1. Examples include: K-Means Clustering, Hierarchical Clustering, DBSCAN, Principal Component Analysis, Association Rule Learning. Scikit-learn provides unsupervised methods for tasks including clustering, dimensionality reduction, nearest-neighbor analysis, and related pattern-discovery problems.
6.5 Clustering and Dimensionality Reduction
1. Clustering: Clustering groups similar observations. Examples: customer segmentation, document grouping, product grouping. Dimensionality Reduction: Dimensionality reduction reduces the number of variables while preserving useful information. Example: reducing 50 related features into a smaller number of components. Principal Component Analysis, or PCA, is a common dimensionality-reduction technique.

7. Reinforcement Learning

7.1 🎮 What is Reinforcement Learning?
1. In reinforcement learning, an agent learns by interacting with an environment. The agent: observes the current situation, chooses an action, receives a reward or penalty, updates its decision strategy, repeats the process. The goal is to learn a strategy that maximizes cumulative reward.
7.2 Important Reinforcement-Learning Terms
1. Agent: The learner or decision-maker. Environment: The world in which the agent acts. State: The current situation observed by the agent. Action: A decision made by the agent. Reward: Feedback received after an action. Policy: The agent's strategy for choosing actions.
7.3 Reinforcement-Learning Examples
1. Robot Navigation: reward for reaching the exit, penalty for collisions, goal: learn an effective path. Game Playing: reward for winning, penalty for losing, goal: learn actions that improve long-term performance. 📌 Reinforcement learning learns from interaction and feedback rather than from fixed labeled examples.

8. Comparing the Main Learning Types

8.1 Quick Comparison
1. A comparison of the three main learning types:

8.2 Selection Questions
1. Ask: Do I have known outputs for past observations? Consider: Supervised learning Do I only have features and want to discover patterns? Consider: Unsupervised learning Does an agent need to learn through actions and rewards? Consider: Reinforcement learning

9. Important Machine-Learning Terms

9.1 🧩 Feature
1. A feature is an input variable used by a model. Examples: age, income, house size, number of rooms. Features are commonly represented using: X
9.2 Label or Target
1. The label, also called the target, is the output that a supervised model tries to predict. Examples: house price, diabetes status, customer churn, salary. The label is commonly represented using: y 📌 Features are the inputs. The label is the desired output.
9.3 Observation
1. An observation is one row or individual example in a dataset. Example: one customer, one house, one patient, one transaction.
9.4 Model
1. A model is a learned mathematical representation of patterns in data. It receives features and produces an output. Features → Model → Prediction
9.5 Training
1. Training is the process in which a model learns patterns from training data. During training, the model adjusts its internal parameters to improve its predictions.
9.6 Prediction
1. A prediction is the output produced by a trained model for an input. Example: Predicted salary = 75,000 Predicted class = Diabetes
9.7 Evaluation
1. Evaluation is the process of measuring how well a trained model performs. The model should be evaluated using data that was not used to fit it. Model evaluation measures prediction quality on unseen data and includes different metrics for classification, regression, and clustering.
9.8 Parameter
1. A parameter is a value learned by the model during training. Example: Coefficients learned by Linear Regression
9.9 Hyperparameter
1. A hyperparameter is a setting chosen before or during the model-development process rather than learned directly from the training data. Examples: Number of neighbors in KNN Maximum depth of a Decision Tree Number of clusters in K-Means Hyperparameter tuning is part of model selection and evaluation workflows.

10. Training and Testing Data

10.1 🧪 Why Split the Data?
1. If we evaluate a model using the same data it learned from, the result may be overly optimistic. We therefore commonly divide the data into: training data, testing data.
10.2 Training Set
1. The training set is used to fit the model. Model learns from training data
10.3 Testing Set
1. The testing set is used to evaluate how well the trained model performs on unseen observations. Model is evaluated on testing data 📌 Training data teaches the model. Testing data checks whether the model generalizes.
10.4 Validation Data
1. A validation set may be used to: compare different models, choose hyperparameters, make development decisions. A simple conceptual split is: Training set → Learn the model Validation set → Choose settings Testing set → Final evaluation Cross-validation and hyperparameter tuning are standard parts of model selection and evaluation.

11. Underfitting and Overfitting

11.1 📉 What is Underfitting?
1. Underfitting occurs when a model is too simple to capture the important patterns in the data. It usually performs poorly on: training data, testing data. Example: Using a straight line to represent a strongly curved relationship. 📌 Underfitting is commonly associated with high bias.
11.2 📈 What is Overfitting?
1. Overfitting occurs when a model learns the training data too closely, including noise or unimportant details. It may perform: very well on training data, poorly on unseen data. 📌 Overfitting means the model memorizes training details but does not generalize well.
11.3 Balanced Model
1. A balanced model captures meaningful patterns without learning unnecessary noise. The goal is not simply to achieve the best training performance. The goal is to perform well on new, unseen data.
11.4 Bias and Variance
1. Bias: Bias is error caused by assumptions that are too simple. High bias is often associated with underfitting. Variance: Variance describes how sensitive a model is to changes in its training data. High variance is often associated with overfitting. Bias-Variance Tradeoff: The bias-variance tradeoff describes the balance between: a model that is too simple, a model that is too sensitive to training data. 📌 A useful model balances learning enough structure without memorizing noise.

12. Data Preparation for Machine Learning

12.1 🧹 Clean the Data
1. Before training a model, we may need to handle: missing values, duplicate rows, incorrect data types, inconsistent categories, outliers, invalid values.
12.2 Feature Encoding
1. Many machine-learning algorithms require numeric input. Categorical values may need to be converted into numbers. Example: City = Kathmandu, Pokhara, Biratnagar may be encoded into numeric columns. This process is called: Encoding
12.3 Feature Scaling
1. Feature scaling places numeric variables on comparable scales. Example: Age: 18 to 80 Salary: 20,000 to 500,000 Some algorithms may be heavily influenced by the larger numeric scale. Common scaling approaches include: standardization, normalization. 📌 Scaling is important for some algorithms, but not every algorithm requires it.
12.4 Avoiding Data Leakage
1. Data leakage occurs when information unavailable at prediction time improperly influences model training or evaluation. A common prevention rule is: split the data before learning preprocessing values from it. For example, a scaler should learn its values from the training data and then apply the same transformation to validation and testing data. Pipelines can organize preprocessing and modeling into a consistent workflow.

13. Model Evaluation

13.1 📏 Why Evaluate a Model?
1. Model evaluation checks whether a model performs well on unseen data. A model that performs well only on its training data may not be useful in practice. Scikit-learn provides separate metric families for classification, regression, clustering, and other model types.
13.2 Classification Metrics
1. Common classification metrics include: accuracy, precision, recall, F1-score, confusion matrix, ROC-AUC.
13.3 Regression Metrics
1. Common regression metrics include: Mean Absolute Error, Mean Squared Error, Root Mean Squared Error, R² score. 📌 The correct evaluation metric depends on the task and the real-world cost of different errors.

14. Steps for Solving Machine-Learning Problems

14.1 Define the Problem
1. Clearly identify: the prediction or discovery goal, the target variable, if one exists, the type of machine-learning task, how success will be measured.
14.2 Collect and Understand the Data
1. Determine: what each row represents, what each column represents, where the data came from, whether the data represents the real problem.
14.3 Explore and Clean the Data
1. Check for: missing values, duplicates, outliers, incorrect types, inconsistent values, unbalanced classes.
14.4 Separate Features and Target
1. For supervised learning: X = df[feature_columns] y = df[target_column] For unsupervised learning: X = df[feature_columns]
14.5 Split the Data
1. Create separate datasets for training and evaluation. Example: training data, testing data. Validation or cross-validation may also be used during model development.
14.6 Preprocess the Features
1. Possible steps include: filling missing values, encoding categorical variables, scaling numeric variables, selecting useful features. Preprocessing should be learned using the training data.
14.7 Choose and Train a Model
1. Select an algorithm appropriate for the task. Examples: Regression → Linear Regression Classification → Logistic Regression Clustering → K-Means Train the model using the training data.
14.8 Evaluate the Model
1. Evaluate the model using suitable metrics and data that was not used to fit it. Compare multiple models when appropriate.
14.9 Improve the Model
1. Possible improvements include: better data cleaning, more useful features, different algorithms, hyperparameter tuning, cross-validation, addressing underfitting or overfitting. These are standard areas within model selection and evaluation.
14.10 Deploy and Monitor
1. After evaluation, an appropriate model may be deployed for real use. Monitoring may check whether: prediction quality changes, input data changes, errors increase, retraining becomes necessary. 📌 Machine learning is an iterative process. Evaluation and improvement may occur repeatedly.

15. Final Machine-Learning Summary

📌 Note
1. The goal of machine learning is not merely to fit past data. It is to learn patterns that remain useful when the model encounters new data.