AI Powered Data Analytics
Led a 7-member team to build a machine learning model predicting learner participation across 8,527 historical records on the Excelerate platform. Using Orange Data Mining, trained and compared three classifiers Logisti
Executive Summary
Led a 7-member team to build a machine learning model predicting learner participation across 8,527 historical records on the Excelerate platform. Using Orange Data Mining, trained and compared three classifiers Logistic Regression, Decision Tree, and Random Forest through a documented three-stage refinement process. The final Random Forest model achieved 93.6% accuracy and 0.980 AUC with 10-fold stratified cross-validation. Generated five data-backed recommendations for program managers, including the finding that 44.2% of opportunities lacked confirmed start dates predicting ne
2The Problem
Excelerate's platform hosted 8,558 learner signups across 23 learning opportunities in 71 countries. The organization needed to identify which opportunities would attract active participation to allocate resources effectively. Without predictive capability, resource allocation relied on intuition rather than evidence.
The dataset contained significant quality issues: 44.3% missing start dates, 1,562 corrupted datetime records, 37+ institution name variants, and no existing predictive model baseline. The team had four weeks to move from raw data to actionable predictions all using no-code Orange Data Mining tools.
Operating Constraints
4-week fixed deadline with weekly deliverable gates. No-code Orange Data Mining tools only no Python, R, or custom scripts. Distributed international team spanning multiple time zones. Dataset had 44.3% missing start dates and 1,562 corrupted datetime records requiring cleaning before any modeling could begin.
3Objective
Build a predictive ML model that classifies opportunities by participation likelihood within 4 weeks, using only no-code Orange Data Mining tools, producing at minimum 90% accuracy on 10-fold stratified cross-validation, and delivering five data-backed executive recommendations.
4Approach & Method
Tools Used
Orange Data Mining (primary), Google Sheets, Google Looker Studio
Why This Approach
Chose Random Forest over Logistic Regression and Decision Tree because it consistently achieved the highest AUC (0.980) across all three refinement stages. Selected Duration Bucket as the primary feature because Information Gain analysis (0.751) showed it dominated all other predictors. Prioritized the mandatory start-date recommendation because it addressed 44.2% of all records with a single operational change, making it the highest-leverage intervention available.
5Execution
6Evidence & Artifacts
7Outcome & Results
- •Final Random Forest model: 93.6% accuracy, 93.7% precision, 93.6% F1, 0.980 AUC. Full-dataset scoring accuracy: 95.3%. Five actionable recommendations delivered, each backed by model-derived evidence and tied to specific platform metrics.
- •The mandatory start-date recommendation alone targets 44.2% of all records with a single operational change. The model provides a reusable prediction framework for all future Excelerate opportunities.
Key Results
- •Final Random Forest model: 93.6% accuracy, 93.7% precision, 93.6% F1, 0.980 AUC. Full-dataset scoring accuracy: 95.3%. Five actionable recommendations delivered, each backed by model-derived evidence and tied to specific platform metrics.
Impact
- •The mandatory start-date recommendation alone targets 44.2% of all records with a single operational change. The model provides a reusable prediction framework for all future Excelerate opportunities.
8What I Learned
- •Feature engineering specifically Duration Bucket creation proved more consequential than algorithm selection. The "No Date" finding demonstrated that data quality interventions can be higher-impact than model tuning. Leading a distributed team through a structured ML workflow within a compressed timeline required balancing technical depth with delivery cadence.