E-commerce Product Classification

This project focuses on classifying product descriptions from an e-commerce website into four categories: Household, Books, Electronics, and Clothing & Accessories. The classification model uses two machine learning algorithms, SVM and Random Forest, and two text representation methods, TF-IDF and Word2Vec.

Table of Contents

• Project Overview
• Project Workflow
• Results
• Conclusion
• License

Project Overview

• Data Source: Product descriptions from an e-commerce website with labels for four categories.
• Text Representation Techniques: TF-IDF and Word2Vec.
• Machine Learning Algorithms: Support Vector Machine (SVM) and Random Forest.
• Performance Metrics: Accuracy, Precision, Recall, and F1 Score.

Project Workflow

1. Data Preprocessing:
   • Clean the data by removing special characters, converting text to lowercase, and performing lemmatization.
   • Remove stopwords to focus on significant words.
2. Text Representation:
   • TF-IDF: Term Frequency-Inverse Document Frequency to represent text as numerical vectors.
   • Word2Vec (CBOW): Continuous Bag of Words to capture the semantic similarity between words.
3. Modeling and Hyperparameter Tuning:
   • Algorithms: Apply SVM and Random Forest with hyperparameter tuning for each text representation method.
   • Tune at least two hyperparameters for each algorithm to improve performance.
4. Evaluation and Comparison:
   • Evaluate model performance on test data using Accuracy, Precision, Recall, and F1 Score.
   • Compare results across different models and text representation techniques to determine the most effective approach.

Results

Text Representation	Algorithm	Hyperparameters	Accuracy	Precision	Recall	F1 Score
TF-IDF (Default)	SVM	c = 1.0, kernel = rbf, gamma = scale	0.9547	0.9553	0.9547	0.9547
	SVM (Tuning 1)	c = 0.5, kernel = linear, gamma = auto	0.9544	0.9546	0.9544	0.9543
	SVM (Tuning 2)	c = 1.5, kernel = poly, gamma = scale	0.8346	0.8783	0.8346	0.8357
	Random Forest	n_estimators = 100, max_depth = None, min_samples_split = 2	0.9345	0.9338	0.9345	0.9344
	Random Forest (Tuning 1)	n_estimators = 150, max_depth = 20, min_samples_split = 5	0.8140	0.8630	0.8140	0.8140
	Random Forest (Tuning 2)	n_estimators = 50, max_depth = 50, min_samples_split = 3	0.9218	0.9260	0.9218	0.9216
TF-IDF (Tuned) min_df = 5, max_df = 0.95	SVM	c = 1.0, kernel = rbf, gamma = scale	0.9551	0.9556	0.9551	0.9551
	Random Forest	n_estimators = 100, max_depth = None, min_samples_split = 2	0.9361	0.9376	0.9361	0.9360
Word2Vec (CBOW)	SVM	c = 1.0, kernel = rbf, gamma = scale	0.9290	0.9292	0.9290	0.9289
	SVM (Tuning 1)	c = 1.5, kernel = linear, gamma = scale	0.9298	0.9298	0.9298	0.9297
	SVM (Tuning 2)	c = 0.5, kernel = poly, gamma = auto	0.8544	0.8567	0.8544	0.8526
	Random Forest	n_estimators = 100, max_depth = None, min_samples_split = 2	0.9405	0.9406	0.9405	0.9405
	Random Forest (Tuning 1)	n_estimators = 75, max_depth = 45, min_samples_split = 4	0.9425	0.9426	0.9425	0.9424
	Random Forest (Tuning 2)	n_estimators = 200, max_depth = 25, min_samples_split = 5	0.9406	0.9406	0.9406	0.9406

Notes:

• Hyperparameter tuning was performed to optimize the models for higher accuracy and F1 scores.
• The SVM model with TF-IDF (Tuned) and Random Forest with Word2Vec (CBOW) achieved the best performance overall.

Conclusion

The results clearly demonstrate that the combination of TF-IDF with a tuned SVM model achieves the best performance across all metrics, including accuracy, precision, recall, and F1 score. Other combinations, such as Random Forest with Word2Vec or even TF-IDF, consistently lagged behind, with their highest scores reaching only around 0.92 compared to 0.95 achieved by the tuned SVM model. These findings highlight the importance of fine-tuning hyperparameters and selecting robust text representation techniques like TF-IDF for achieving superior results in e-commerce product classification. This approach proves particularly effective for tasks requiring both accuracy and consistency in categorization.

License

This project is licensed under the MIT License - see the LICENSE file for details.