MoodLens AI

MoodLens AI is an explainable NLP application that detects emotions in text using a fine-tuned RoBERTa classifier and shows why predictions are made through SHAP token-level explanations.

Live Application: https://moodlens-ai-igo7kb8nrjmvmrcqhti9xy.streamlit.app/

Developer: Dev Dharmesh Patel
GitHub: https://github.com/devpatel0005
Email: devdpatel0005@gmail.com

---

Project Summary

This project solves a practical problem in applied NLP: most emotion classifiers return only a label, but do not explain decision logic. MoodLens AI provides both prediction and interpretability in one product-style interface.

For each input sentence, the app provides:

• Predicted emotion class
• Confidence score
• Per-class confidence distribution chart
• SHAP Summary Plot for token importance
• SHAP Force Plot for token contribution direction

Emotion classes supported:

• sadness
• joy
• love
• anger
• fear
• surprise

---

Project Highlights

• Built and deployed an end-to-end explainable AI web app for emotion detection.
• Integrated Hugging Face Transformers with Streamlit for real-time NLP inference.
• Implemented SHAP explainability in production-style UI using predicted-class token attributions.
• Added confidence distribution visualization for transparent model behavior.
• Added robust class-label mapping logic to normalize LABEL_X outputs into human-readable class names.
• Optimized inference flow using cached model loading and optional CUDA support.
• Deployed the app publicly and maintained dependency reliability for cloud runtime.

---

Model and Explainability Details

Base Model

• Architecture: RoBERTa sequence classifier
• Framework: Hugging Face Transformers + PyTorch
• Output: six-class emotion probabilities

Explainability Layer

• Method: SHAP (SHapley Additive exPlanations)
• Explainer input: text classification pipeline
• Scope: explanations are generated for the predicted class
• Visual outputs:
  • Summary bar plot: token impact magnitude
  • Force plot: token push-up and push-down effects from baseline to final score

Inference Pipeline

1. User input is collected through Streamlit UI.
2. Pipeline performs tokenization and model inference.
3. Probabilities are ranked and displayed as chart + metric.
4. SHAP explainer computes token-level contributions.
5. Summary and force plots are rendered in the app.

---

Model Performance (Test Set)

Emotion	Precision	Recall	F1-Score
Sadness	0.97	0.98	0.97
Joy	0.95	0.96	0.95
Love	0.92	0.90	0.91
Anger	0.94	0.94	0.94
Fear	0.91	0.91	0.91
Surprise	0.89	0.85	0.87

---

Dataset Summary

The emotion detection model is trained and evaluated on a comprehensive Twitter emotion dataset:

Dataset Overview:

• Total Samples: 16,000 rows
• Total Features: 3 columns
  • text: Raw Twitter text/emotion statements
  • label: Actual emotion class (0-5)
  • predicted_label: Model's prediction

---

Tech Stack

• Python
• Streamlit
• PyTorch
• Hugging Face Transformers
• SHAP
• Plotly
• Pandas
• NumPy
• Matplotlib
• scikit-learn

---

Repository Structure

• frontend/app.py: Streamlit application logic, inference workflow, SHAP rendering
• models/emotion_model/: trained model artifacts and tokenizer files
• notebooks/Emotion_detection.ipynb: model development and evaluation workflow
• notebooks/shap.ipynb: SHAP experimentation and interpretation workflow
• datasets/emotion_predictions_full.csv: project data artifact
• requirements.txt: deployment and local dependency list
• docs/: documentation assets

---

Local Setup and Run

1) Clone Repository

git clone https://github.com/devpatel0005/MoodLens-AI.git
cd MoodLens-AI

2) Install Dependencies

python -m pip install -r requirements.txt

3) Run Streamlit App

streamlit run frontend/app.py

---

Deployment Notes

• Live deployment is running on Streamlit Cloud.
• Ensure all required libraries are listed in requirements.txt before deploy.
• Current dependency set includes plotly and torchvision to prevent cloud import errors.
• Model files are loaded from local project paths in the deployed environment.

---

Skills Demonstrated

• End-to-end machine learning productization
• NLP model integration and inference engineering
• Explainable AI implementation in user-facing applications
• Data visualization for model confidence communication
• Cloud deployment debugging and dependency management
• UI/UX iteration for ML applications

---

Limitations

• Predictions may be less reliable on very short or ambiguous text.
• SHAP explanations are local and input-specific, not full causal proofs.
• Domain shift can affect generalization on unseen language styles.

---

Planned Improvements

• Confidence thresholding and low-confidence fallback messaging
• Batch inference support for multiple texts
• Additional explainability views for class comparison
• Monitoring and lightweight analytics for production usage

---

License

This repository is for educational and portfolio purposes.