🌟 YıldızSezar: Turkish E-Commerce Review Classifier

This repository contains the official implementation of YıldızSezar, my B.Sc. Thesis Project. The research has been peer-reviewed and published in the International Journal of 3D Printing Technologies and Digital Industry.

🔗 Read the Full Paper Here (A Star Rating-Based Approach in BERT-Based Sentiment Analysis of Customer Feedback)
🤖 Try the Live Model on Hugging Face Spaces
📊 Explore the Dataset on Hugging Face

🎥 Hugging Face Space Demo

📌 Project Overview

The objective of YıldızSezar is to automatically predict 1-to-5 star ratings directly from morphologically complex Turkish customer reviews.

Real-world e-commerce data often suffers from extreme class imbalance (heavily skewed towards 5-star and 1-star reviews). To combat this, I engineered a scalable synthetic data generation pipeline using LLaMA-8B-DPO, generating over 900,000 synthetic review samples for minority classes.

The final ConvBERT-based classifier achieved an accuracy of 89.96% and a macro F1-score of 0.799, significantly outperforming traditional ML baselines.

🚀 Key Features & Production Readiness

• Advanced Deep Learning Models: Fine-tuned ConvBERT, DistilBERT, and ELECTRA.
• Synthetic Data Augmentation: Solved extreme class imbalance using open-source LLMs.
• Production Performance Testing: Implemented scripts to measure Latency (ms) and Throughput (QPS) for production environments.
• Reliability & Calibration: Analyzed Expected Calibration Error (ECE) and generated reliability diagrams to ensure the model's confidence scores are trustworthy.

📊 Model Evaluation & Calibration

(The charts below demonstrate the model's robustness across different classes and its predictive confidence).

📂 Repository Structure

• /data_processing: Scripts for data cleaning, HTML unescaping, and dataset splitting.
• /training: PyTorch/Transformers training loops, including Distributed Data Parallel (DDP).
• /inference: Scripts for basic model loading and terminal inferences.
• /evaluation: Advanced evaluation scripts (Latency, Throughput, Expected Calibration Error).
• /web_app: Flask-based Web Interface for local demonstrations.

💻 Quick Start (Python)

You can easily integrate the model using the Hugging Face transformers library:

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_id = "ilkayO/yildizsezar-convbert"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id)

review = "Biraz gecikmeli aktarıyor görüntüyü ama onun dışında fiyat performans ürünü."
inputs = tokenizer(review, return_tensors="pt", truncation=True, max_length=256)

with torch.no_grad():
    outputs = model(**inputs)
    predicted_class_id = torch.argmax(outputs.logits, dim=-1).item()

print(f"Predicted Star Rating: {predicted_class_id + 1} Stars")

⚖️ License

This project utilizes a split-licensing strategy:

• Source Code: GNU GPL v3.0
• Model Weights: Apache 2.0
• Dataset: CC BY-NC 4.0