Customer Review Insights Dashboard

A Python-based analytical tool that uses RoBERTa for sentiment analysis, K-Means clustering for automated product categorization, and BART for abstractive review summarization.

Features

• Sentiment Analysis: Classifies customer reviews (Positive / Neutral / Negative) using a fine-tuned RoBERTa model.
• Product Clustering: Groups products into logical categories using sentence embeddings + K-Means, then labels each cluster with GPT-4o-mini.
• Visualization: Interactive dashboard (Chart.js) showing sentiment distribution, category rankings, and sentiment trends over time.
• AI Market Reports (notebook-only, not wired into the UI): Generates "best"/"worst" product summaries using facebook/bart-large-cnn, producing cohesive write-ups instead of raw quote lists.

Project Structure

app.py                  # Flask application / API routes
services/
  data_processor.py      # Upload validation & cleaning
  ml_engine.py           # Sentiment, clustering, and category labeling
models/sentiment_model/  # Fine-tuned RoBERTa checkpoint (not included, see below)
notebooks/               # Model training & experimentation
  classifier_full_dataset.ipynb
  clustering.ipynb
  review_summarization.ipynb
templates/, static/       # Dashboard UI

Installation & Setup

1. Clone the repository

   git clone https://github.com/keirialaa/customer_reviews.git
   cd customer_reviews

2. Create a virtual environment and install dependencies

   python -m venv .venv
   source .venv/bin/activate  # Windows: .venv\Scripts\activate
   pip install -r requirements.txt        # runtime only
   # or, to also run the notebooks / tests:
   pip install -r requirements-dev.txt

3. Configure environment variables

   Copy the example file and fill in your own keys:

   cp .env.example .env

   Variable	Required	Purpose
   FLASK_SECRET_KEY	Yes	Flask session signing key. Any random string works locally.
   OPENAI_API_KEY	Yes	Used by GPT-4o-mini to generate cluster category labels.
   HF_TOKEN	Only for the notebooks	Hugging Face token used when downloading/pushing models in notebooks/.

4. Provide the sentiment model

   models/sentiment_model/ is not committed to the repo (the checkpoint is ~500MB). To get it locally, either:

   • Run notebooks/classifier_full_dataset.ipynb end-to-end, which fine-tunes RoBERTa and saves the checkpoint to models/sentiment_model/, or
   • Point MODEL_PATH in services/ml_engine.py at a Hugging Face Hub model ID instead of a local path.

5. Run the application

   export FLASK_APP=app.py
   flask run

   Then open http://127.0.0.1:5000, upload a CSV with at least product_name and review_text columns, and view the dashboard.

Running tests

pip install -r requirements-dev.txt
pytest

License

MIT — see LICENSE.