LMU CS Capstone Project: Synthetic Donor Dataset & Multimodal Fusion Model for Fundraising Analytics

🎯 Project Overview

Fundraising and philanthropy research suffers from a scarcity of open-source datasets—privacy concerns and donor confidentiality make real-world data almost impossible to share. This project addresses that gap by generating a synthetic donor dataset (500K records, 1.8M dense relationships) expressly designed for experimenting with machine learning (ML) and deep learning (DL) approaches in advancement analytics.

The dataset includes contact reports, gift histories, event attendance, household linkages, and graph-based influence signals. On top of the data layer, we implement a novel multimodal fusion model that unifies tabular, sequential, network, and text-derived features through a deep learning pipeline (MLP + LSTM + attention + fusion layers). The result is an end-to-end DL pipeline for exploring how modern AI techniques can elevate prospect prioritization, revenue forecasting, and gift officer workflows.

Key Innovations

• Open source synthetic fundraising dataset with rich, interleaved donor behaviors.
• Multimodal fusion architecture combining tabular encoders, sequence models, graph/network embeddings, and text aggregates with cross-modal attention.
• Temporal validation (1980–2025) to mimic real deployment scenarios for 2025 predictions.
• Streamlit dashboard that surfaces KPIs, feature insights, and business impact for practitioner review.

📊 Dataset Snapshot

• Total Donors: 500,000
• Dense Relationships: 1,805,144
• Giving History Records: 3,836,541
• Event Attendance: 156,065
• Contact Reports: 329,299
• Relationship Mix: Professional (500K), Alumni (401K), Geographic (300K), Giving (300K), Activity (200K), Family (93K), Social (10K)

🗂️ Project Structure

LMUCapstoneProject/
├── data/
│   └── synthetic_donor_dataset_500k_dense/      # Core synthetic corpus
│       ├── donors.csv / donor_database.db       # Tabular + SQL formats
│       ├── dense_relationships.(csv|parquet)    # Network edges
│       ├── giving_history.csv                   # Longitudinal gifts
│       └── parts/                               # Generation checkpoints
├── dashboard/                                   # Streamlit app + assets
│   ├── app.py                                   # Main entry point
│   ├── architecture_diagram.html                # Visual overview (used in README)
│   └── pages/ / components/ / models/           # Modular UI + metrics
├── docs/                                        # Deep dives + guides
├── scripts/                                     # Dataset generation utilities
├── src/                                         # ML/DL training pipelines
├── models/                                      # Saved checkpoints
└── visualizations/                              # Plots, analyses, figures

🧠 Architecture Diagram

The full multimodal training flow (data generation → feature engineering → fusion model → evaluation → dashboard deployment) is captured in dashboard/architecture_diagram.html. Open that file in a browser to view the interactive layered diagram referenced throughout the documentation.

🖥️ Streamlit Dashboard

To access the Streamlit dashboard: https://fictitiousuniversity.streamlit.app/

🚀 Quick Start

1. Generate / refresh the dataset

python scripts/generate_enhanced_500k_dataset_with_dense_relationships.py

2. Explore via SQLite (recommended)

import sqlite3, pandas as pd
conn = sqlite3.connect('data/synthetic_donor_dataset_500k_dense/donor_database.db')
donors = pd.read_sql_query("SELECT * FROM donors LIMIT 10", conn)
relationships = pd.read_sql_query("""
    SELECT * FROM relationships
    WHERE Relationship_Category = 'Alumni'
    LIMIT 10
""", conn)

3. Work with CSV/Parquet extracts

import pandas as pd
donors = pd.read_csv('data/synthetic_donor_dataset_500k_dense/donors.csv')
relationships = pd.read_csv('data/synthetic_donor_dataset_500k_dense/dense_relationships.csv')

4. Launch the dashboard

streamlit run dashboard/app.py

🔧 Technical Highlights

Memory + Performance

• Incremental disk writes and chunked processing prevent OOM on standard laptops.
• Vectorized NumPy/Pandas flow yields 10–100× faster generation.
• SQLite backend ships with 24 tuned indexes → complex network joins in ~3 seconds.

Data Fidelity

• Referential integrity enforced across donors, gifts, relationships, and events.
• Graph density calibrated (≈0.0014%) to mirror enterprise advancement CRMs.
• QA pipeline plus resumable checkpoints for long-running jobs.

Modeling Approach

• Feature store covers RFM, engagement streaks, network centrality, capacity indicators, and synthetic contact report stats.
• Multimodal pipeline (src/models/train_will_give_again.py) fuses 60+ engineered features with sequential gift histories, network embeddings, and SVD-based text signals.
• Training splits: 1980–2023 (train), 2024 (validation), 2025 (test target) with AdamW + BCE-with-logits + ReduceLROnPlateau + batch size 2048 + hidden dim 256.

📚 Documentation & Support

• docs/TRAINING_PIPELINE_GUIDE.md – Training + evaluation steps
• docs/INTERPRETABILITY_GUIDE.md – Explaining model decisions
• docs/OPTIMIZATION_GUIDE.md – Performance + scaling tips
• examples/ – Quick notebooks and scripts to jump-start analysis

For questions, explore the docs/ folder or open an issue/PR with reproducible steps. This repository is intentionally open so other institutions can build upon the synthetic dataset and multimodal modeling blueprint.

---

👤 Credits

Created by
Danielle Brown
Loyola Marymount University
M.S. in Computer Science Senior Capstone Project
2025