Charisma-Predictor is a multimodal AI pipeline developed as part of a Master's research project in Artificial Intelligence at Maastricht University. It received a final grade of 8.5 / 10, based on methodology, implementation, and evaluation.

This repository contains the parts I personally developed: the video model, the fusion logic, and the charisma scoring and visualization.

• Video: facial landmark tracking via MediaPipe + five sequence models (CNN, LSTM, GRU, Transformer, TCN)
• Ensemble: video outputs fused via weighted averaging (best MAE: 0.1189)
• Fusion: weighted averaging and Multi-Channel Weighted Fusion (MCWF)
• Output: Big Five personality prediction + charisma score ∈ [0, 1], classified into five levels
• Optimization: early stopping, learning rate scheduling (ReduceLROnPlateau)
• Visuals: leadership score distribution, confusion matrices
• Achieved up to 92.45% accuracy on Big Five prediction via fusion (MCWF)

Leadership Suitability Distribution:

The fusion output shows high alignment between predicted and true leadership scores, confirming the effectiveness of cross-modal aggregation.

Fusion Model Personality Accuracy (MCWF):

This repository reflects my direct contributions to the project:

• Developed the full video model pipeline: feature extraction, five-model ensemble, evaluation
• Designed and implemented fusion logic, including MCWF and weighted fusion strategies
• Created the charisma scoring system using normalized Big Five trait correlations
• Generated final evaluation metrics, plots, and analysis outputs

charisma-predictor/
├── video_model/            # Facial landmark + sequence model ensemble
├── fusion/                 # Fusion logic (weighted avg, MCWF)
├── figures/                # Output plots (confusion matrices, histograms)
├── results/                # Personality predictions + charisma scores
├── text_and_audio/         # External references to group members' models
│   └── README.md
├── report_links/           # Final report (PDF)
│   └── README.md
├── LICENSE 
├── README.md               # You're reading it
└── requirements.txt

While this repo focuses on my implementation, the final fusion model also incorporated audio and text inputs from teammates:

• 🔊 Audio model (AST)
• 📄 Text model (BERT)

This repository includes the training code for the fusion model. To train the Multi-Channel Weighted Fusion (MCWF) model:

pip install -r requirements.txt
cd fusion
python train_fusion.py

• First Impressions Dataset – 10,000 annotated video clips

Model Workflow Overview:

• Video: MediaPipe landmark sequences → five-model ensemble (CNN, LSTM, GRU, Transformer, TCN)
• Audio: AST + Random Forest (team contribution)
• Text: BERT-based personality estimation (team contribution)
• Fusion: Average, weighted, MCWF
• Output: Big Five scores → 0–1 charisma score → five-class suitability label