Md Rifat Ahmmad Rashid, Associate Professor, Dept.

of CSE
East West University, Bangladesh

Lecture 12: Machine Learning

Overview of Machine Learning
Machine learning (ML) is a subset of artificial intelligence (AI) that enables systems to learn from
data, identify patterns, and make decisions with minimal human intervention. It is grounded in the
idea that systems can automatically improve their performance on a task with experience, typically
through training on a large amount of data. Machine learning algorithms can be broadly categorized
into three types: supervised learning, unsupervised learning, and reinforcement learning.

• Supervised Learning: The algorithm learns from labeled data, meaning the input data is
associated with known outcomes or labels.
• Unsupervised Learning: The algorithm learns from unlabeled data and tries to uncover
hidden patterns without any explicit output labels.
• Reinforcement Learning: The algorithm learns through trial and error by interacting with an
environment and receiving feedback through rewards or penalties.

Machine Learning Workflow

The machine learning workflow involves several key stages:

1. Problem Definition: Understanding the problem at hand and deciding whether machine
Learning is the right approach.
2. Data Collection: Gathering relevant data that represents the problem space.
3. Data Preprocessing: Cleaning, transforming, and normalizing data to ensure it is suitable
for the machine learning model.
4. Feature Engineering: Selecting or extracting relevant features (variables) that will be used
by the model.
5. Model Selection: Choosing the appropriate machine learning algorithm(s) for the task.
6. Model Training: Feeding the model with training data and allowing it to learn patterns.
7. Model Evaluation: Assessing the model's performance using test data and appropriate
metrics.
8. Model Tuning: Optimizing model parameters (hyperparameters) to improve performance.
9. Deployment: Implementing the model in a real-world environment to make predictions.
10. Monitoring and Maintenance: Continuously monitoring model performance and updating
it as needed based on new data.

1|Pa ge
 Md Rifat Ahmmad Rashid, Associate Professor, Dept. of CSE
East West University, Bangladesh

2|Pa ge
 Md Rifat Ahmmad Rashid, Associate Professor, Dept. of CSE
East West University, Bangladesh

Fine-Tuning a Model

Fine-tuning is a process of improving model performance by adjusting model parameters,

architecture, or training processes. Here’s how students can fine-tune their models:

1. Adjust Hyperparameters:
• Learning Rate: Decrease the learning rate if the model's loss is fluctuating wildly.
Increase it slightly if the model is learning too slowly.
• Batch Size: A larger batch size can stabilize gradient updates but requires more memory.
A smaller batch size can lead to noisier updates but often helps in generalization.

2. Regularization:
• L2 Regularization (Ridge): Helps to reduce overfitting by adding a penalty to large
weights.
• Dropout (for deep learning): Randomly "drops" neurons during training to prevent the
network from becoming too reliant on any one node, thus reducing overfitting.

3. Early Stopping:

• Stop training when the validation loss stops improving. This can prevent overfitting, as
the model will stop when it starts performing poorly on the validation set, even if the
training loss is decreasing.

4. Optimizer Adjustment:
• For deep learning models, using optimizers like Adam or RMSprop often improves
performance over traditional SGD (Stochastic Gradient Descent). Try different
optimizers or adjust the learning rate schedule (e.g., step decay, exponential decay).

Bias-Variance Tradeoff

The bias-variance tradeoff is a fundamental concept in machine learning that describes the
balance between two sources of errors:

Bias: Refers to the error introduced by approximating the real-world problem too simply. High
bias occurs when the model is too simple and underfits the data.

Example: A linear model trying to capture a non-linear relationship.

Variance: Refers to the error introduced by the model’s sensitivity to small fluctuations in the
training data. High variance occurs when the model is too complex and overfits the data.

Example: A deep neural network with many layers memorizing training data.

3|Pa ge
 Md Rifat Ahmmad Rashid, Associate Professor, Dept. of CSE
East West University, Bangladesh

Goal: The goal is to find the "sweet spot" where the model complexity is just right, minimizing
both bias and variance.

• Underfitting: High bias and low variance — model too simple.

• Overfitting: Low bias and high variance — model too complex.

Strategies to Manage Bias-Variance Tradeoff:

For High Bias (Underfitting):

• Use a more complex model (e.g., from linear regression to polynomial regression).
• Increase the number of features or add more layers to neural networks.
For High Variance (Overfitting):
• Use regularization techniques (L1, L2 regularization).
• Increase the size of the training dataset.
• Reduce the model complexity (e.g., reduce the depth of a decision tree).

Epochs and Iterations

In deep learning, understanding the relationship between epochs and iterations is critical for
setting up proper training.

Epoch: One epoch means that the model has seen the entire training data once.
Iteration: One iteration means that the model has updated its weights once after processing a
batch of data.

Choosing the Ideal Number of Epochs:

• The number of epochs depends on the dataset's size and complexity.
• Too few epochs may lead to underfitting, while too many may lead to overfitting.
• Use early stopping to automatically halt training when the validation error stops
improving.

Typical Rule of Thumb:

• For small datasets: 50–100 epochs.
• For large datasets (e.g., ImageNet): Several hundred epochs may be necessary.

Monitoring Training:
• Track training and validation loss across epochs.
• Stop training when the validation loss starts to increase, indicating overfitting.

4|Pa ge
 Md Rifat Ahmmad Rashid, Associate Professor, Dept. of CSE
East West University, Bangladesh

Data Splitting and Cross-Validation

Data Splitting: Overview and Ideal Ratios

Data splitting involves partitioning the dataset into different subsets to train, validate, and test the
model. The goal is to evaluate the model's performance reliably while maximizing the use of
available data.

Ideal Data Splits for ML Problems

1. Training Set:
o The largest portion (typically 60-80%) of the data is used for training the model.
o The model learns patterns, adjusts weights, and minimizes error on this subset.
2. Validation Set:
o 10-20% of the data is reserved for validation.
o It is used to fine-tune the model’s hyperparameters, check for overfitting, and
evaluate the model's performance during training.
3. Test Set:
o The remaining 10-20% of the data is set aside as a test set, which is used to assess
the final performance of the model.
o This set is kept entirely separate from the training and validation processes to
provide an unbiased evaluation of the model's accuracy.

Data Splits for Specific Types of Data

• Image Data:
o When dealing with image datasets, it's common to split data into training (70-
80%), validation (10-15%), and test (10-15%) sets. For smaller image datasets,
additional techniques like data augmentation are often applied to increase the
diversity of the training set.
o For object detection and segmentation tasks, maintaining a balanced distribution
of classes and object scales across the splits is crucial.
• NLP Data:
o NLP datasets are often split in the same ratios as typical machine learning
problems (e.g., 70% train, 15% validation, 15% test). However, special attention
is needed to maintain a balanced representation of different classes, languages, or
document types (e.g., sentiment classes or entity types) across all splits.
o Additionally, for sequence data, ensuring that similar sequences are not
overrepresented in the training set compared to validation and test sets is
essential.

5|Pa ge
 Md Rifat Ahmmad Rashid, Associate Professor, Dept. of CSE
East West University, Bangladesh

Image source: https://www.linkedin.com/pulse/data-splitting-strategies-machine-learning-swapnil-sharma/

Cross-Validation

Cross-validation is a technique used to assess how well a model will generalize to an

independent dataset. By training and validating a model multiple times with different data
partitions, cross-validation provides a more accurate and reliable measure of model performance.

K-Fold Cross-Validation:

1. Concept: The dataset is divided into K subsets (or folds). The model is trained on K-1
folds and validated on the remaining fold. This process is repeated K times, with each
fold serving as the validation set once. The performance metric is averaged across all K
iterations.
o When to Use:
▪ Ideal for small to medium-sized datasets where it's crucial to maximize the
use of all available data.
▪ Suitable for most types of machine learning problems (e.g., tabular data,
image data, and NLP).

6|Pa ge
 Md Rifat Ahmmad Rashid, Associate Professor, Dept. of CSE
East West University, Bangladesh

o Benefit: Reduces bias and variance by ensuring that every data point is used for
both training and validation. It provides a more reliable measure of model
performance, particularly for small datasets where every observation is valuable.

Image Source: https://medium.com/@ompramod9921/cross-validation-623620ff84c2

When to Use Cross-Validation

• Small to Medium-Sized Datasets: When you have limited data, cross-validation

maximizes the use of your dataset by ensuring every point is used for both training and
validation.
• Model Selection and Hyperparameter Tuning: Cross-validation helps compare
different models or hyperparameter settings reliably, as it provides consistent
performance estimates across multiple data splits.
• Algorithm Evaluation: For models like decision trees, logistic regression, or SVMs,
cross-validation is a suitable method to validate their performance, especially when the
dataset is not large.

When Not to Use Cross Validation

• Large Datasets: When working with large datasets (e.g., millions of observations), a
simple train-validation-test split is often sufficient, as the dataset is already large enough
to provide reliable estimates.
• Computationally Expensive Models: For deep learning models or other complex
algorithms that require significant training time, cross-validation may not be practical due

7|Pa ge
 Md Rifat Ahmmad Rashid, Associate Professor, Dept. of CSE
East West University, Bangladesh

to the need to train multiple models. In such cases, using a separate validation set or
employing alternative validation strategies like early stopping is more efficient.

Monitoring and Evaluation

Once the model is trained, it’s critical to evaluate its performance:
• Use the test set: Ensure that the test data has never been used during training or
validation.
Plot learning curves: Monitor both the training and validation error throughout the

1. Evaluation Metrics for Regression

Regression tasks involve predicting continuous numerical values. Evaluation metrics for regression
measure the difference between the predicted and actual values.

1. Mean Absolute Error (MAE):

o Definition: The average of the absolute differences between predicted and actual
values.

o Formula:

o Interpretation: Lower values indicate better performance.

2. Mean Squared Error (MSE):

o Definition: The average of the squared differences between predicted and actual
values.

o Formula:

o Interpretation: Penalizes larger errors more heavily than MAE. Lower values are
better.

3. Root Mean Squared Error (RMSE):

o Definition: The square root of the MSE. It represents the standard deviation of the
residuals (prediction errors).

8|Pa ge
 Md Rifat Ahmmad Rashid, Associate Professor, Dept. of CSE
East West University, Bangladesh

o Formula:

o Interpretation: Lower RMSE indicates a better fit, making it more interpretable in

the original unit of the target variable.

4. R-Squared (Coefficient of Determination):

o Definition: Measures the proportion of the variance in the dependent variable that
is predictable from the independent variables.

o Interpretation: Values range from 0 to 1, with higher values indicating better model
performance.

5. Adjusted R-Squared:

o Definition: A modified version of R-Squared that accounts for the number of

predictors in the model. It penalizes for adding irrelevant variables.

o Interpretation: Useful when comparing models with different numbers of

predictors.

6. Mean Absolute Percentage Error (MAPE):

o Definition: The average of the absolute percentage errors between predicted and
actual values.

o Formula:

o Interpretation: Expressed as a percentage; lower values indicate better

performance.

7. Explained Variance Score:

o Definition: Measures the proportion of variance explained by the model, similar to

R-squared but can be negative.

9|Pa ge
 Md Rifat Ahmmad Rashid, Associate Professor, Dept. of CSE
East West University, Bangladesh

o Range: [-∞, 1], where 1 indicates perfect prediction and lower values indicate worse
performance.

2. Evaluation Metrics for Classification

Classification tasks involve predicting categorical outcomes. Metrics for classification are designed
to evaluate the accuracy and robustness of such predictions.

1. Accuracy:

o Definition: The ratio of correctly predicted instances to the total number of

instances.

o Formula:

o Interpretation: Suitable when the dataset is balanced, but less informative for
imbalanced datasets.

2. Precision:

o Definition: The ratio of true positive predictions to the total positive predictions.

o Formula:

o Interpretation: Indicates how many positive predictions are actually correct; useful
in cases where false positives are costly.

3. Recall (Sensitivity):

o Definition: The ratio of true positive predictions to the total actual positives.

o Formula:

o Interpretation: Indicates how many actual positives the model captures; crucial in
scenarios where false negatives are costly.

4. F1 Score:

o Definition: The harmonic mean of precision and recall, providing a balance

between the two.

o Formula:

10 | P a g e
 Md Rifat Ahmmad Rashid, Associate Professor, Dept. of CSE
East West University, Bangladesh

o Interpretation: Useful for imbalanced datasets where accuracy alone is not

sufficient.

5. Area Under the ROC Curve (AUC-ROC):

o Definition: Measures the model’s ability to distinguish between classes by plotting

the True Positive Rate (TPR) against the False Positive Rate (FPR).

o Range: [0, 1], where 1 indicates perfect separation and 0.5 indicates no separation.

o Interpretation: Higher values indicate better model performance.

Image source: https://x.com/mdancho84/status/1752019456664973641

11 | P a g e
 Md Rifat Ahmmad Rashid, Associate Professor, Dept. of CSE
East West University, Bangladesh

Image source: https://towardsdatascience.com/understanding-the-roc-curve-in-three-visual-steps-

795b1399481c

6. Area Under the Precision-Recall Curve (AUC-PR):

o Definition: A plot that shows the trade-off between precision and recall for different
thresholds.

o Interpretation: Especially informative for imbalanced datasets, where the AUC-

ROC may not provide enough insight.

7. Logarithmic Loss (Log Loss):

o Definition: Measures the uncertainty of the predictions. It penalizes incorrect

predictions with higher penalties for confident but incorrect predictions.

o Formula:

12 | P a g e
 Md Rifat Ahmmad Rashid, Associate Professor, Dept. of CSE
East West University, Bangladesh

o Interpretation: Lower values indicate better performance.

What is Supervised Learning?

In supervised learning, the dataset consists of inputs (features) and their corresponding outputs
(labels or target values). The model is trained using this labeled dataset, and the objective is to
learn a mapping function from the input to the output. The learning process involves minimizing the
error between the model’s predictions and the actual outputs.

• Input (Features): Variables or attributes of the data that are used as input for the model
(e.g., age, height, income).

• Output (Label/Target): The variable that the model aims to predict (e.g., house price,
classification category).

Image source: https://medium.com/@metehankozan/supervised-and-unsupervised-learning-an-intuitive-

approach-cd8f8f64b644

13 | P a g e
 Md Rifat Ahmmad Rashid, Associate Professor, Dept. of CSE
East West University, Bangladesh

Types of Supervised Learning Problems

Supervised learning problems can be broadly categorized into two types: regression and
classification.

1. Regression:

o Definition: Regression involves predicting continuous numerical values. The output

variable is quantitative and can take any real value.

o Examples:

▪ Predicting house prices based on features like square footage, location, and
number of bedrooms.

▪ Forecasting temperature changes over time based on historical weather

data.

o Common Algorithms:

▪ Linear Regression

▪ Decision Trees for Regression

▪ Random Forest Regressor

▪ Support Vector Regressor (SVR)

▪ Neural Networks (for continuous outputs)

14 | P a g e
 Md Rifat Ahmmad Rashid, Associate Professor, Dept. of CSE
East West University, Bangladesh
Image source: https://www.geeksforgeeks.org/ml-linear-regression/

2. Classification:

o Definition: Classification involves predicting discrete class labels or categories.

The output variable is categorical and takes one of a limited set of values (e.g.,
binary or multi-class).

o Examples:

▪ Classifying email as spam or not spam.

▪ Predicting whether a tumor is benign or malignant based on medical

images.

o Common Algorithms:

▪ Logistic Regression

▪ Decision Trees

▪ Random Forest Classifier

▪ Support Vector Machine (SVM)

▪ k-Nearest Neighbors (k-NN)

▪ Naive Bayes

▪ Neural Networks (for categorical outputs)

The Supervised Learning Process

The process of supervised learning generally involves the following steps:

1. Data Collection:

o Gather a dataset that includes both inputs (features) and corresponding outputs
(labels). The dataset must be labeled accurately for the model to learn effectively.

2. Data Preprocessing:

o Cleaning: Remove or handle missing values, outliers, and inconsistencies.

o Normalization/Standardization: Scale features to a uniform range (e.g., 0 to 1) or a

standard distribution (mean = 0, variance = 1).

o Encoding: Convert categorical features into numerical format using techniques like
one-hot encoding or label encoding.

o Splitting Data: Divide the dataset into training, validation, and test sets to evaluate
model performance accurately.

3. Model Selection:

15 | P a g e
 Md Rifat Ahmmad Rashid, Associate Professor, Dept. of CSE
East West University, Bangladesh

o Choose a suitable algorithm based on the type of problem (regression or

classification), the size of the dataset, and the complexity of the relationships within
the data.

4. Model Training:

o The model is trained on the training set using a chosen algorithm. The model
iteratively updates its parameters to minimize the error (loss function) between the
predicted and actual outputs.

5. Model Evaluation:

o Evaluate the model’s performance using the validation set. Performance metrics
vary based on the type of task:

▪ For regression: Mean Absolute Error (MAE), Mean Squared Error (MSE), R-
squared.

▪ For classification: Accuracy, Precision, Recall, F1 score, Area Under the

ROC Curve (AUC-ROC).

6. Hyperparameter Tuning:

o Adjust the model’s hyperparameters (e.g., learning rate, depth of a decision tree) to
optimize performance. This is often done using techniques like grid search or
random search with cross-validation.

7. Model Testing:

o The final model is tested on the test set, which is unseen during training and
validation, to assess its generalization ability.

Key Concepts in Supervised Learning

1. Training and Test Sets:

o Training Set: Used to train the model. It represents the majority of the data.

o Test Set: Used to evaluate the final model’s performance. It provides an unbiased
estimate of how the model will perform on new data.

2. Loss Function:

o A function that measures the error between the predicted output and the actual
output. The objective of training is to minimize this error.

o Common Loss Functions:

▪ For regression: Mean Squared Error (MSE), Mean Absolute Error (MAE).

▪ For classification: Cross-entropy loss, Hinge loss.

3. Optimization Algorithms:

16 | P a g e
 Md Rifat Ahmmad Rashid, Associate Professor, Dept. of CSE
East West University, Bangladesh

o Methods used to minimize the loss function and update the model’s parameters.
These algorithms determine how the model learns from the data.

o Examples:

▪ Gradient Descent

▪ Stochastic Gradient Descent (SGD)

▪ Adam Optimizer (Adaptive Moment Estimation)

4. Bias-Variance Tradeoff:

o Bias: The error introduced by approximating a real-world problem with a simplified

model. High bias leads to underfitting.

o Variance: The model’s sensitivity to small fluctuations in the training data. High
variance leads to overfitting.

o The goal is to find a balance between bias and variance to achieve optimal model
performance.

5. Cross-Validation:

o A technique used to evaluate model performance and reduce overfitting by splitting

the dataset into multiple folds. K-fold cross-validation is a common method where
the dataset is divided into K subsets, and the model is trained and validated K times.

o When to Use: Suitable for small to medium-sized datasets to maximize data

utilization.

Common Algorithms in Supervised Learning

1. Linear Regression:

o Use Case: Predicting continuous outcomes based on linear relationships between

features.

o Example: Estimating house prices.

2. Logistic Regression:

o Use Case: Binary classification tasks, such as spam detection.

o Example: Predicting whether an email is spam or not.

3. Decision Trees:

o Use Case: Both regression and classification tasks.

o Example: Classifying customers based on their likelihood of purchasing a product.

4. Support Vector Machines (SVM):

17 | P a g e
 Md Rifat Ahmmad Rashid, Associate Professor, Dept. of CSE
East West University, Bangladesh

o Use Case: Classification tasks, particularly with high-dimensional data.

o Example: Text classification in NLP.

5. k-Nearest Neighbors (k-NN):

o Use Case: Simple and intuitive classification and regression tasks based on
similarity measures.

o Example: Recommending products based on similar user preferences.

6. Naive Bayes:

o Use Case: Classification tasks, especially in text classification where

independence assumptions can hold.

o Example: Sentiment analysis in social media posts.

7. Ensemble Methods:

o Random Forest: Combines multiple decision trees for improved accuracy and
stability.

o Gradient Boosting (e.g., XGBoost): Sequentially trains models to correct the errors
of previous ones.

o Use Case: Effective in both regression and classification, particularly for complex
datasets.

Advantages and Disadvantages of Supervised Learning

Advantages:

1. Accuracy: Supervised learning models can achieve high accuracy if there is sufficient
labeled data and the model is well-tuned.

2. Interpretability: Some models (e.g., linear regression, decision trees) are easy to interpret
and understand.

3. Versatility: It can be applied to a wide range of problems, including regression, binary

classification, and multi-class classification.

Disadvantages:

1. Dependence on Labeled Data: Supervised learning requires a significant amount of

labeled data, which can be expensive and time-consuming to obtain.

2. Overfitting: Models may perform well on the training data but poorly on new data if they
overfit. Regularization techniques and cross-validation are used to mitigate this risk.

3. Scalability: For large datasets or complex tasks (e.g., deep learning models), training can
be computationally intensive and time-consuming.

18 | P a g e