Topic Classifier using TF-IDF and Linear SVM

Overview

This project implements a multi-class topic classification system that predicts the topic of a document based on its textual content. The solution uses classical machine learning techniques by combining TF-IDF feature extraction with a Linear Support Vector Machine (Linear SVM).

The project was developed as part of the IIT Gandhinagar Technical Assistant assignment. The complete machine learning pipeline includes dataset inspection, exploratory data analysis, text preprocessing, feature engineering, model development, experimentation, evaluation, error analysis, and inference.

The original dataset contains approximately 10 million documents (4 GB). During development and experimentation, a representative subset of 122,880 documents was used to enable efficient training and evaluation on CPU hardware.

---

Final Model Performance

Metric	Value
Final Model	Linear Support Vector Machine (Linear SVM)
Feature Extraction	TF-IDF (Word Unigrams + Bigrams)
Development Dataset	122,880 Documents
Original Dataset	10 Million Documents
Number of Classes	24
Vocabulary Size	100,000
Accuracy	92.06%

---

Features

• Multi-class topic classification for 24 predefined categories
• TF-IDF feature extraction using word unigrams and bigrams
• Linear Support Vector Machine (Linear SVM) classifier
• Logistic Regression baseline implementation
• Balanced Logistic Regression implementation
• Hybrid TF-IDF + Linear SVM implementation
• Automatic evaluation using Accuracy, Precision, Recall, and F1-score
• Confusion matrix generation
• Misclassification analysis
• Saved trained model using Joblib
• Saved TF-IDF vectorizer for inference
• Interactive command-line inference script

---

Project Structure

Topic-Classifier/
│
├── data/
│   └── dataset_10.parquet
│
├── figures/
│   └── confusion_matrix.png
│
├── models/
│   ├── baseline_lr.joblib
│   ├── balanced_lr.joblib
│   ├── hybrid_svm.joblib
│   ├── linear_svm.joblib
│   └── tfidf_vectorizer.joblib
│
├── outputs/
│   ├── baseline_metrics.csv
│   ├── balanced_lr_metrics.csv
│   ├── hybrid_svm_metrics.csv
│   ├── linear_svm_metrics.csv
│   └── misclassified_examples.csv
│
├── reports/
│   └── report.pdf
│
├── src/
│   ├── config.py
│   ├── loader.py
│   ├── preprocessing.py
│   ├── feature_engineering.py
│   ├── feature_engineering_hybrid.py
│   ├── model.py
│   ├── model_balanced_lr.py
│   ├── model_svm.py
│   ├── evaluate.py
│   ├── train_baseline.py
│   ├── train_balanced_lr.py
│   ├── train_hybrid_svm.py
│   ├── train_svm.py
│   ├── inference.py
│   ├── error_analysis.py
│   ├── plot_confusion_matrix.py
│   ├── inspect_dataset.py
│   ├── eda.py
│   └── data_profiler.py
│
├── README.md
├── requirements.txt
└── .gitignore

---

Dataset

The dataset was provided as part of the IIT Gandhinagar Technical Assistant assignment.

Dataset Information

Property	Value
Format	Parquet
File Name	dataset_10.parquet
Size	Approximately 4 GB
Total Rows	10,000,000
Columns	DATA, TOPIC

Column Description

Column	Description
DATA	Input text document
TOPIC	Target topic label

Place the provided dataset inside the data/ directory before training.

Note: The dataset is not included in this repository because it was provided as part of the assignment.

---

Installation

Clone the repository

git clone https://github.com/ayushgit9090/Topic-Classifier.git
cd Topic-Classifier

Create a virtual environment

python -m venv .venv

Activate the virtual environment

Windows

.venv\Scripts\activate

Linux / macOS

source .venv/bin/activate

Install dependencies

pip install -r requirements.txt

---

Training

Train the final Linear SVM model using

python src/train_svm.py

The training pipeline performs:

• Dataset loading
• Text preprocessing
• TF-IDF feature extraction
• Train/Test split
• Linear SVM training
• Model evaluation
• Model serialization
• TF-IDF vectorizer serialization

Saved files:

models/
├── linear_svm.joblib
└── tfidf_vectorizer.joblib

---

Inference

Run the trained classifier

python src/inference.py

Example

Input

Virat Kohli scored a century in today's cricket match.

Output

Predicted Topic:
sports_and_fitness

Another example

Input

Python programming using NumPy and pandas.

Output

Predicted Topic:
software_development

---

Input / Output Schema

Input

A single text document.

Example

Python programming using NumPy and pandas.

Output

One predicted topic label.

Example

software_development

---

Experimental Results

Model	Accuracy
Logistic Regression	84.90%
Balanced Logistic Regression	71.50%
Hybrid TF-IDF + Linear SVM	88.10%
Linear SVM	92.06%

The Linear SVM achieved the highest overall accuracy and was selected as the final model.

---

Reproducibility

The project ensures reproducibility by:

• Using a fixed random seed for train/test splitting.
• Applying deterministic preprocessing.
• Using a fixed TF-IDF configuration.
• Saving both the trained model and fitted TF-IDF vectorizer.
• Providing a standalone inference script.
• Maintaining a modular project structure.

---

Future Improvements

Possible future enhancements include:

• Training on the complete 10 million document dataset.
• Incremental learning for large-scale training.
• Better handling of minority classes.
• Advanced feature engineering.
• Distributed training for faster experimentation.

---

Author

Ayush Das

2026