4.1.2 Model Building in Machine Learning

Introduction

Model building in machine learning involves several key steps to develop a system that can learn from data and make predictions or decisions. The textbook explains that this is a systematic process — from collecting and cleaning data to selecting an algorithm, training the model, evaluating its performance, and finally deploying it for real-world use.

• Collection of Data from various sources.
• Data Cleaning and Feature Engineering.
• Model Building for selecting the correct ML algorithm.
• Evaluate Model.
• Model Deployment.

Definition

Model Building in Machine Learning is the process of creating a mathematical or computational model that represents relationships between variables in data to make predictions or gain insights.

In simple words: Model building is the step-by-step process of teaching a computer to make predictions from data.

Main Idea

The core idea is that a machine learning model is not created in one step — it is built through a series of stages. At first, data is collected and preprocessed to ensure it is clean and suitable for analysis. This might involve normalizing values, handling missing data, and encoding categorical variables.

After data preprocessing, a suitable machine learning algorithm is chosen based on the problem at hand — whether it is classification, regression, clustering, or another type. The algorithm is then trained on the dataset, adjusting its parameters to minimize error and improve accuracy.

During training, the model's performance is evaluated using a separate validation set of data to ensure its accuracy about new unseen data. Once the model is optimized, it is tested on a final test dataset to assess its effectiveness. If the results are satisfactory, the model can be deployed for real-world use; otherwise, it may require further improvements in code.

Core Concept

The Five Steps of Model Building (Fig. 4.3)

Step 1: Collection of Data Data is collected from various sources. This is the foundation of the entire process — without data, there is no model. The quality and quantity of the data directly affect how well the model will perform.

• Handling missing data.
• Removing duplicates.
• Correcting errors.
• Encoding categorical variables.

Feature engineering is the process of selecting and transforming the most relevant data features to improve a model's performance.

• Classification: Sorting data into categories.
• Regression: Predicting continuous values.
• Clustering: Grouping similar data points.
• Other types.

The algorithm is then trained on the dataset, adjusting its parameters to minimize error and improve accuracy.

Step 4: Evaluate Model The model's performance is evaluated using a separate validation set of data to ensure its accuracy on new, unseen data. Once the model is optimized, it is tested on a final test dataset to assess its effectiveness.

If the results are satisfactory, the model can be deployed. Otherwise, it may require further improvements.

Step 5: Model Deployment If the model performs well, it is deployed for real-world use. This means the model is integrated into a system where it can make predictions or decisions on new data.

Throughout this process, every time the model reads data and calculates results, the algorithm keeps on learning. Repeated refinement and evaluation help in achieving a model that performs well and meets the specific needs of the application.

Key Terms in Model Building (from the book)

The textbook uses a real-life problem to explain key terms: developing a model to identify or differentiate images of dogs, cats, and horses. A reasonable collection of images of all three animals is needed. Each term is explained with reference to this example.

Feature Engineering Feature engineering is the process of selecting and transforming the most relevant data features to improve a model's performance.

Example from the book: If you want to predict that a given image is of a dog, cat, or horse, you have to make a list of common and distinct features of each animal — i.e., size, color, appearance, shape of eyes, ears, hair, nose, etc. The model will keep data about all possible variations of features of these animals, so that a sitting horse cannot be confused with the image of a big size standing dog.

Test-Train Split A test-train split divides data into two parts: training data (main portion) and test data (smaller portion).

Example from the book: In the animal image example, the textbook decided to make a test-train split of seventy and thirty percent (70% training, 30% test).

• If too much data is provided at the time of training, the model will overfit.
• If too little data is provided for training, the model will underfit.

Model Training Model training is the key feature in machine learning. It helps the model make decisions that are more accurate and meaningful. Without proper model training, the machine is unable to decide about the new data.

The textbook identifies three commonly used machine learning models:

1. Supervised Learning: The model is trained using labeled data, where the correct output is provided for each input. - Example from the book: Linear Regression, used to predict continuous values such as prices or temperatures.

2. Unsupervised Learning: The data is unlabeled, and the model tries to find patterns or groupings on its own. - Example from the book: K-Means Clustering, which groups similar data points together, such as in customer segmentation.

3. Reinforcement Learning: The model learns by interacting with an environment, receiving feedback in the form of rewards or penalties. - Example from the book: Q-Learning, often used in game playing or robotic navigation where the model learns to make decisions to maximize its rewards.

Overfitting Overfitting happens when a machine learning model learns the training data too well, including its noise and irrelevant details. This makes the model perform very well on the training data but poorly on new, unseen data because it's too specific to the training set.

Underfitting Underfitting occurs when a model is too simple and fails to capture the underlying patterns in the data. As a result, it performs poorly both on the training data and on new data, as it hasn't learned enough from the available information.

Model Assessment Model assessment evaluates a model's performance using parameters like accuracy, precision, and recall.

Example from the book: Suppose 100 new images of cats, dogs, and horses (not part of the train and test data) are provided. If the model predicts the correct results for 80 images and for the remaining 20 images it predicts wrong labels (e.g., it gives a horse label for an image of a dog), then the accuracy of the machine learning model is 80%.

Accuracy Accuracy is the percentage of total predictions that are correct. It tells how often the model is right overall.

Accuracy = Correct PredictionsTotal Predictions

Precision Precision measures how many of the predictions the model made for a certain class (e.g., "positive" cases) are actually correct. It's about the quality of the positive predictions.

Precision = True PositivesTrue Positives + False Positives

Intermediate Level

The five steps of model building are interconnected:

• Feature Engineering: Selecting and transforming relevant features.
• Test-Train Split: Dividing data into training (70%) and test (30%) sets.
• Overfitting: Model learns training data too well; performs poorly on new data.
• Underfitting: Model is too simple; performs poorly on all data.
• Accuracy: Percentage of correct predictions.
• Precision: Quality of positive predictions.

Important Points

• - Model Building Steps: Data collection, data cleaning & feature engineering, model building, evaluation, deployment.
• - Feature Engineering: Selecting and transforming relevant features.
• - Test-Train Split: Divides data into training (main portion) and test (smaller portion). Common split: 70% training, 30% test.
• - Model Training: Supervised (labeled data), Unsupervised (unlabeled data), Reinforcement (rewards/penalties).
• - Overfitting: Model learns training data too well; performs poorly on new data.
• - Underfitting: Model is too simple; performs poorly on all data.
• - Model Assessment: Evaluates performance using accuracy, precision, recall.
• - Accuracy: Correct predictions / Total predictions.
• - Precision: True positives / (True positives + False positives).

Remember

• - 5 Steps: Collect → Clean → Build → Evaluate → Deploy.
• - Feature Engineering = Select/transform features.
• - Test-Train Split = 70% training, 30% test.
• - Supervised = Labeled data.
• - Unsupervised = Unlabeled data.
• - Reinforcement = Rewards/penalties.
• - Overfitting = Too much training.
• - Underfitting = Too little training.
• - Accuracy = Correct / Total.
• - **Precision** = TP / (TP + FP).

Common Mistakes

• 1. Skipping data cleaning — Dirty data leads to poor models.
• 2. Using the same data for training and testing — This leads to overfitting and overly optimistic results.
• 3. Choosing the wrong algorithm — The algorithm must match the problem type (classification, regression, clustering).
• 4. Confusing overfitting with underfitting — Overfitting is too much training; underfitting is too little.
• 5. Thinking accuracy is always the best metric — Accuracy can be misleading for imbalanced data; precision and recall are also important.
• 6. Ignoring feature engineering — The quality of features directly affects model performance.

Quick Revision

• Model Building in Machine Learning: The process of creating a mathematical or computational model that represents relationships between variables in data to make predictions or gain insights.
• Five Steps:
• Key Terms:
• Feature Engineering: Selecting and transforming the most relevant data features.
• Test-Train Split: Divides data into training (main portion) and test (smaller portion). Common split: 70% training, 30% test.
• Model Training: Supervised (labeled data), Unsupervised (unlabeled data), Reinforcement (rewards/penalties).
• Overfitting: Model learns training data too well; performs poorly on new data.
• Underfitting: Model is too simple; performs poorly on all data.
• Model Assessment: Evaluates performance using accuracy, precision, recall.
• Accuracy: Correct predictions / Total predictions.
• Precision: True positives / (True positives + False positives).

60-Second Revision

• 5 Steps: Collect → Clean → Build → Evaluate → Deploy.
• Feature Engineering = Select/transform features.
• Test-Train Split = 70% training, 30% test.
• Supervised = Labeled data.
• Unsupervised = Unlabeled data.
• Reinforcement = Rewards/penalties.
• Overfitting = Too much training.
• Underfitting = Too little training.
• Accuracy = Correct / Total.
• **Precision** = TP / (TP + FP).
• Warning: Choose the test-train split wisely to avoid overfitting or underfitting.

Frequently Asked Questions (FAQs)

Q1. What is model building in machine learning? Answer: Model building in machine learning is the process of creating a mathematical or computational model that represents relationships between variables in data to make predictions or gain insights. It involves several key steps: data collection, data cleaning and feature engineering, model building, evaluation, and deployment.

• Collection of Data: Gather raw data from various sources.
• Data Cleaning and Feature Engineering: Handle missing data, encode variables, select relevant features.
• Model Building: Choose and train the algorithm.
• Evaluate Model: Assess performance using validation and test sets.
• Model Deployment: Integrate the model into real-world systems.

Q3. What is feature engineering? Answer: Feature engineering is the process of selecting and transforming the most relevant data features to improve a model's performance. For example, if you want to predict whether an image is of a dog, cat, or horse, you would list common and distinct features of each animal — size, color, appearance, shape of eyes, ears, hair, nose, etc.

Q4. What is a test-train split? Answer: A test-train split divides data into two parts: training data (main portion) and test data (smaller portion). The textbook uses a 70% training, 30% test split. This ratio should be chosen wisely — too much training data leads to overfitting, and too little leads to underfitting.

• Supervised Learning: Trained using labeled data (e.g., Linear Regression).
• Unsupervised Learning: Finds patterns in unlabeled data (e.g., K-Means Clustering).
• Reinforcement Learning: Learns through rewards and penalties (e.g., Q-Learning).

Q6. What is overfitting? Answer: Overfitting happens when a machine learning model learns the training data too well, including its noise and irrelevant details. This makes the model perform very well on the training data but poorly on new, unseen data because it's too specific to the training set.

Q7. What is underfitting? Answer: Underfitting occurs when a model is too simple and fails to capture the underlying patterns in the data. As a result, it performs poorly both on the training data and on new data, as it hasn't learned enough from the available information.

• Q8. What is accuracy?
• Answer: Accuracy is the percentage of total predictions that are correct. It tells how often the model is right overall.
• Accuracy = Correct PredictionsTotal Predictions

Q9. What is precision? Answer: Precision measures how many of the predictions the model made for a certain class (e.g., "positive" cases) are actually correct. It's about the quality of the positive predictions. Precision = True PositivesTrue Positives + False Positives

Q10. What is model assessment? Answer: Model assessment evaluates a model's performance using parameters like accuracy, precision, and recall. For example, if a model predicts the correct results for 80 out of 100 new images, the accuracy is 80%.

Q11. Why is feature engineering important? Answer: Feature engineering is important because the quality of features directly affects model performance. By selecting and transforming the most relevant features, you can improve the model's ability to learn patterns and make accurate predictions.

Q12. What is the difference between supervised and unsupervised learning? Answer: Supervised learning uses labeled data, where the correct output is provided for each input. Unsupervised learning uses unlabeled data, and the model tries to find patterns or groupings on its own.

Practice Questions

Short Response Questions (SRQs)

• Define model building in machine learning.
• List the five steps of model building.
• What is feature engineering? Give an example.
• What is a test-train split? Why is the ratio important?
• Define overfitting and underfitting.

Extended Response Questions (ERQs)

• Discuss the five steps of model building in machine learning in detail. Why is each step important?
• Explain the key terms in model building (feature engineering, test-train split, overfitting, underfitting, accuracy, precision) with examples.
• Analyze the difference between supervised, unsupervised, and reinforcement learning. Provide examples of each.
• You are building a machine learning model to classify images of animals. Describe how you would apply the five steps of model building, including feature engineering and test-train split.
• Explain the importance of accuracy and precision in model assessment. Why is it important to use multiple metrics?

Multiple Choice Questions (MCQs)

• Q1. What is the first step in model building?
• A. Data cleaning
• B. Data collection
• C. Model evaluation
• D. Model deployment
• Correct Answer: B
• Explanation: The first step is data collection.

• Q2. What is feature engineering?
• A. Collecting raw data
• B. Transforming and selecting relevant data features
• C. Testing the model's accuracy
• D. Calculating the significance of the model
• Correct Answer: B
• Explanation: Feature engineering involves selecting and transforming the most relevant data features.

• Q3. What is a test-train split?
• A. Dividing data into training and test portions
• B. Splitting the algorithm into two parts
• C. Dividing the model into two layers
• D. Splitting the output into two categories
• Correct Answer: A
• Explanation: A test-train split divides data into training data (main portion) and test data (smaller portion).

• Q4. What is overfitting?
• A. The model is too simple
• B. The model learns the training data too well, including noise
• C. The model performs well on all data
• D. The model does not learn at all
• Correct Answer: B
• Explanation: Overfitting happens when a model learns the training data too well, including noise and irrelevant details.

• Q5. What is underfitting?
• A. The model is too complex
• B. The model learns the training data too well
• C. The model is too simple and fails to capture patterns
• D. The model performs well on all data
• Correct Answer: C
• Explanation: Underfitting occurs when a model is too simple and fails to capture underlying patterns.

• Q6. What is supervised learning?
• A. Learning from unlabeled data
• B. Learning from labeled data
• C. Learning through rewards and penalties
• D. Learning without data
• Correct Answer: B
• Explanation: Supervised learning uses labeled data, where the correct output is provided for each input.

• Q7. What is unsupervised learning?
• A. Learning from labeled data
• B. Learning from unlabeled data
• C. Learning through rewards and penalties
• D. Learning without data
• Correct Answer: B
• Explanation: Unsupervised learning uses unlabeled data, and the model tries to find patterns or groupings on its own.

• Q8. What is reinforcement learning?
• A. Learning from labeled data
• B. Learning from unlabeled data
• C. Learning through rewards and penalties
• D. Learning without data
• Correct Answer: C
• Explanation: Reinforcement learning involves a model that learns by interacting with an environment, receiving feedback in the form of rewards or penalties.

• Q9. What is the formula for accuracy?
• A. True Positives / (True Positives + False Positives)
• B. Correct Predictions / Total Predictions
• C. Total Predictions / Correct Predictions
• D. False Positives / True Positives
• Correct Answer: B
• Explanation: Accuracy = Correct Predictions / Total Predictions.

• Q10. What is the formula for precision?
• A. Correct Predictions / Total Predictions
• B. True Positives / (True Positives + False Positives)
• C. False Positives / True Positives
• D. Total Predictions / Correct Predictions
• Correct Answer: B
• Explanation: Precision = True Positives / (True Positives + False Positives).

• Q11. Why is the test-train split ratio important?
• A. Too much training data leads to overfitting; too little leads to underfitting
• B. Too much training data leads to underfitting; too little leads to overfitting
• C. The ratio does not matter
• D. The ratio only affects the speed of training
• Correct Answer: A
• Explanation: If too much data is provided at the time of training, the model will overfit. If too little data is provided for training, the model will underfit.

• Q12. What is model assessment?
• A. Evaluating a model's performance using accuracy, precision, and recall
• B. Collecting data
• C. Cleaning data
• D. Deploying the model
• Correct Answer: A
• Explanation: Model assessment evaluates a model's performance using parameters like accuracy, precision, and recall.