👥 MinneMUDAC 2025 Winners — Predicting Match Longevity for Big Brothers Big Sisters

🎉 We’re thrilled to share our award-winning submission for MinneMUDAC 2025!
Our team, We Push to Prod, placed 2nd overall for building an end-to-end machine learning pipeline to predict how long a volunteer mentorship match will last at Big Brothers Big Sisters Twin Cities.

---

🎯 Problem Statement

Most mentorship matches end early, limiting long-term impact.
How can we predict match longevity and identify at-risk matches before they close?

---

📊 Dataset Overview

• Structured data: Match metadata, demographic attributes, census-level community features
• Unstructured data: Match support notes, closure descriptions
• Source: Provided by Big Brothers Big Sisters Twin Cities

---

🧠 Our Approach

1. 📌 Exploratory Analysis

• Matches with >18 year age gap lasted ~7 months longer
• Ethnicity/gender alignment boosted match longevity by ~5 months
• Community-based programs yielded longer matches vs. school-based

2. 💬 NLP & Sentiment Engineering

• Extracted BERT embeddings from match support notes
• Computed sentiment and emotional tone slopes over last 3 check-ins
• Mapped closure notes to root reasons using custom keyword clustering

3. 🔍 Predictive Modeling

• Target: Match length in months
• Baseline: Linear regression (RMSE: 17 months)
• Final model: XGBoost
  • RMSE: 10.2 months
  • Features used:
    • Demographics (age diff, ethnicity match, gender match)
    • BERT sentiment/emotion slopes
    • Shared interests
    • Match type (community vs. school)

---

🔁 Solution Pipeline

graph TD;
    A[Raw Data] --> B[EDA & Cleaning]
    B --> C[BERT Embeddings + Sentiment Slopes]
    B --> D[Structured Feature Engineering]
    C & D --> E[XGBoost Model]
    E --> F[Match Length Prediction + Closure Risk Flags]

Loading

🧭 Revamped Engagement Flowcharts for Coordinators

To improve early detection of at-risk matches, we developed a structured, decision-tree–based system for monthly check-ins. These trees guide match coordinators in identifying key issues and mapping them to root causes and intervention strategies.

---

🚫 Mentor Decision Tree – Not Currently Meeting

---

✅ Mentor Decision Tree – Currently Meeting

---

👥 Mentee Decision Tree

---

💡 Why This Matters

• Converts vague updates into structured data
• Links dropdown responses to root causes (e.g., time, health, or compatibility issues)
• Enables targeted intervention strategies
• Prevents silent disengagement by surfacing emotional and logistical blockers

---

🧠 Insights & Takeaways

• 📝 Text sentiment over time is a leading indicator of match breakdown.
• 👥 Matches under optimal demographic conditions lasted 7–9 months longer.
• ✅ Our redesigned mentor check-in questionnaire helps identify high-risk closures early.

---

🛠 Tools & Stack

• 🐍 Python, Pandas, XGBoost
• 🤖 BERT (transformer embeddings)
• 🎭 Hartman Emotional Tone model
• 📊 SHAP, Seaborn, Scikit-learn
• 📽 PowerPoint (final slides)

---

🏆 Team: We Push to Prod

• Justin Varghese – Lead Data Scientist (Modeling, Feature Engineering)

---

📬 Contact

Built with ❤️ for social good and data science.
Feel free to reach out to Justin Varghese for collaboration, mentorship, or speaking opportunities.

🔗 GitHub Profile