Text Classifier (Decision Tree)

Overview

This project implements a text classifier in Java using a decision-tree representation. The classifier is trained on labeled text data loaded from CSV files, builds a tree using word-frequency features and numeric thresholds, and can classify new text inputs by traversing the learned tree.

The classifier also supports saving and reloading trained models using a preorder tree format.

This project was completed as part of CSE 123 (Computer Programming) at the University of Washington.

---

Features

• Decision-tree classifier with recursive traversal
• Word-frequency feature extraction
• Threshold-based branching decisions
• CSV-based training and testing data loading
• Save and reload trained classifier trees
• Accuracy evaluation on test datasets

---

Key Concepts Used

• Binary trees
• Recursion
• Object-oriented design
• File I/O
• Feature extraction from text
• Data-driven algorithm design

---

Project Structure

• Classifier.java // Core classifier logic and tree operations
• TextBlock.java // Represents labeled text and word-frequency data
• DataLoader.java // Loads training and testing data
• CsvReader.java // Parses CSV files into usable records
• Client.java // Runs training, testing, and classification

---

How It Works

Text Representation

• Each document is represented as a TextBlock.
• A TextBlock stores:
  • A label (classification category)
  • A mapping of words to occurrence counts
  • The total number of words in the document
• Word probabilities are computed as: count(word) / totalWords

---

Training the Classifier

• The classifier is trained on a list of labeled TextBlock objects.
• Internal decision nodes store:
• A feature word
• A numeric threshold
• During training, examples that are misclassified cause the tree to grow by introducing a new decision node that separates examples based on whether the word probability meets the threshold.

---

Classification

• To classify new text, the classifier:

1. Starts at the root of the tree
2. Evaluates the stored feature against its threshold
3. Recursively traverses left or right
4. Returns the label stored at the leaf node

---

Saving and Loading Models

• The classifier can be saved to a file using a preorder traversal format.
• Each node is written as either:
• A feature/threshold pair (branch node), or
• A label (leaf node)
• The saved file can later be reloaded to reconstruct the exact same tree structure.

---

Example Workflow

1. Load training data from a CSV file
2. Train the classifier on labeled text examples
3. Evaluate accuracy on a separate test dataset
4. Save the trained classifier to disk
5. Reload the classifier and classify new inputs

---

What I Learned

• How decision trees encode conditional logic
• How recursion simplifies tree traversal and reconstruction
• How feature selection and thresholds affect classification behavior
• How to design programs that separate data loading, modeling, and execution

---

Future Improvements

• Support multi-feature splits
• Improve feature selection heuristics
• Visualize the decision tree structure
• Add probabilistic confidence scores to classifications

---

Notes

This project was completed as coursework. All implementation and design decisions are my own.