Semi Supervised Classification

Semi supervised classification is a technique in data mining that uses both labeled and unlabeled data to build a classification model. Usually, only a small portion of the dataset has labels, while the remaining data is unlabeled. The model learns patterns from both types of data to improve classification performance.

• Uses both labeled and unlabeled data.
• Only a small portion of data is labeled.
• Most data remains unlabeled.
• Useful when labeling data is difficult or costly.
• Combines supervised and unsupervised learning

Working

The working of semi supervised classification in data mining can be explained in the following steps:

1. Collect Labeled and Unlabeled Data: Gather a dataset that contains a small amount of labeled data and a large amount of unlabeled data.

Example: A few images of animals with labels and many images without labels.

2. Choose a Classification Algorithm: Select a suitable classification algorithm that can work with both labeled and unlabeled data.

3. Train the Model with Labeled Data: The model first learns patterns and relationships using the available labeled data.

4. Use Unlabeled Data for Learning: The model then uses the unlabeled data to discover additional patterns and improve its understanding.

5. Improve Classification Performance: By combining both types of data, the model becomes better at classifying new or unseen data.

Implementation

Step 1: Import Required Libraries

Import the necessary Python libraries for data handling, machine learning and evaluation.

Step 2: Load the Dataset

• The Iris dataset is loaded from sklearn.
• It contains 150 samples and 3 flower classes.
• Features represent measurements of petals and sepals.

Output:

Step 3: Visualize the Dataset

This visualization shows relationships between features.

• Each color represents a different class.
• It helps understand how the classes are distributed.
• Good separation between classes makes classification easier.

Output:

Step 4: Split the Dataset

• 70% of data is used for training.
• 30% of data is used for testing.
• The model learns from training data and is evaluated on test data.

Step 5: Create Unlabeled Data

Semi supervised learning assumes that most data is unlabeled.

• Some training labels are removed.
• -1 represents unlabeled samples.
• The model will try to infer these labels during training.

Step 6: Visualize Labeled vs Unlabeled Data

This plot shows the distribution of labeled and unlabeled samples.

• Colored points represent labeled data.
• Dark points represent unlabeled data.
• This demonstrates the semi-supervised setting.

Output:

Step 7: Train the Semi Supervised Model

The model learns patterns using:

• Available labeled data
• Structure of unlabeled data

Step 8: Predict the Test Data

The trained model predicts the class of unseen samples.

Step 9: Evaluate Model Accuracy

• Accuracy measures how many predictions are correct.
• Higher accuracy means the classification model performs better.

Output:

Accuracy: 1.0

Step 10: Confusion Matrix Visualization

The confusion matrix shows:

• Correct predictions along the diagonal
• Misclassifications in other cells

Output:

Download full code from here

Applications

• Used in audio data to improve speech recognition systems.
• Since billions of web pages exist, labeling them manually is not practical. It helps automatically classify and organize web content.
• Uses a few labeled texts and many unlabeled ones to classify documents into categories.

Advantages

• Semi supervised classification is relatively simple to understand and implement because it combines ideas from supervised and unsupervised learning.
• Requires only a small amount of labeled data, which reduces the effort and cost involved in manually labeling large datasets.
• Algorithm can still learn useful patterns by using a large amount of unlabeled data along with the available labeled samples.

Disadvantages

• Results of the algorithm may change across different iterations because the model continuously updates labels during training.
• May not effectively capture complex relationships present in large network or graph based datasets.
• Accuracy may not always be high, especially when the unlabeled data contains noise or misleading patterns.