🧠 Big Five Personality Clustering & Classification Project

📌 Project Overview

This project focuses on analyzing and classifying psychological profiles using responses to the Big Five Personality Test.
The methodology integrates unsupervised learning (for cluster discovery) and supervised learning (for classification), followed by an ensemble approach to maximize predictive performance.
Additionally, an interactive Streamlit web application was developed to allow users to take the Big Five test and receive an ML-powered personality profile.

---

📂 Dataset

• Source: OpenPsychometrics Big Five open dataset
• Features: 50 item-level responses (Likert scale)
• Target: Cluster labels derived from Gaussian Mixture Model (GMM) clustering

---

🧩 Unsupervised Learning (GMM)

✅ Clustering Process

• Applied PCA for dimensionality reduction (90% explained variance).
• Used Gaussian Mixture Models (GMM) with full covariance.
• Model selection based on AIC, BIC, and Silhouette Score.
• Optimal number of clusters: 4.

✅ Cluster Profiles

The final clusters were manually labeled based on their OCEAN centroids:

Cluster	Name	Description
0	Reserved	Moderate Openness, Low Agreeableness and Conscientiousness, slightly elevated Neuroticism
1	Striver	Moderate-to-high O, C, E, A, with elevated Neuroticism (driven achievers but emotionally tense)
2	Internalizer	Low O and E, moderate A, very high Neuroticism (introverted and emotionally vulnerable)
3	Balanced	Average O, C, E, A with very low Neuroticism (emotionally stable and well-adjusted)

---

🔥 Supervised Learning

Once clusters were identified, the problem was reframed as a 4-class classification task.

✅ Models Tested

1. XGBoost
2. Multi-Layer Perceptron (MLP) (TensorFlow/Keras)

Both models were trained first with default parameters, then optimized with Optuna for hyperparameter tuning.

---

🎯 Results Summary

Model	Tuning	Accuracy	F1-Score
XGBoost	No tuning	0.889	0.890
XGBoost (Optuna)	✅ Optuna	0.914	0.914
MLP	No tuning	0.922	0.921
MLP (Optuna)	✅ Optuna	0.937	0.937
Soft Voting (XGB + MLP)	✅ Ensemble	0.940	0.940

---

🤝 Ensemble Learning (Soft Voting)

✅ Approach

• Combined the probabilities predicted by XGBoost and MLP using a weighted average.
• Grid search was used to find optimal weights maximizing the F1-score.

✅ Optimal Weights

• XGB weight: 0.3
• MLP weight: 0.7

✅ Performance

• Accuracy: 0.9401
• F1-score: 0.9401

---

🖥️ final_pipeline.py

The script final_pipeline.py is a ready-to-run inference pipeline that:

1. Loads the optimized models (xgb_optuna.pkl and mlp_tuning.h5)
2. Loads the dataset (big5_supervised_dataset.csv) and splits it into train/test
3. Generates predictions using both XGBoost and MLP
4. Applies soft voting with optimal weights
5. Evaluates model performance and plots the confusion matrix
6. Saves predictions and class probabilities for further analysis

---

🌐 Streamlit Web App (big5_app/)

An interactive Streamlit web app was implemented, allowing users to:

• Take the Big Five Personality Test directly in the browser (50 questions)
• Get predictions from the trained ensemble model (XGB + MLP)
• Visualize results via:
  • Class probabilities bar chart
  • Radar chart of OCEAN traits
  • Display of predicted psychological profile

The app was structured to run both locally and can be deployed online (e.g., via Streamlit Cloud or Heroku) without requiring Python installation by the user.

---

✅ Conclusions

• GMM clustering successfully identified 4 distinct psychological profiles.
• MLP Optuna achieved the best individual performance.
• Soft Voting Ensemble further improved accuracy and robustness.
• The addition of the Streamlit app enables interactive personality testing using the trained models.

---

🚀 Future Work

• Extend dataset with cross-cultural samples.
• Explore SHAP for explainability.
• Test additional ensemble methods (stacking, blending).
• Deploy the app to a cloud service for public access.

---

🛠️ Technologies Used

• Python, scikit-learn, Optuna, TensorFlow/Keras, XGBoost, Streamlit
• Matplotlib, Pandas, NumPy

---

👤 Author

Developed by Samuele as part of an advanced machine learning project on psychological profiling.