Learner Personas with Clustering

🎓 Learner Personas

Behavioral learner segmentation using unsupervised machine learning to discover actionable learner personas.

📊 Key Result Card

📸 Persona Snapshot

🎯 Problem Statement

Educational platforms serve thousands of learners, each with highly varied engagement patterns. The goal of this project is not predictive modeling (e.g., predicting churn), but rather discovering meaningful behavioral segments that can support differentiated product strategies.

Clustering is appropriate here because it allows the data to organically reveal natural behavioral groupings (personas) without imposing biased demographic or pre-conceived psychological boundaries.

📂 Dataset

The raw dataset captures the learning behavior of real users.
Importantly, the raw dataset is strictly read-only and is never modified during the pipeline.

• Size: 6,000 learners
• Shape: 18 columns
• Timeframe: 2023-01-01 → 2025-06-19
• Demographics: 8 countries
• Tech context: 3 platforms, 2 subscription types, 2 notification states
• Quality: 93 missing values in avg_session_minutes, 0 duplicate rows, 0 duplicate learner IDs

⚙️ Feature Engineering

We ran explicit experiments comparing StandardScaler (behavior_standard) vs RobustScaler (behavior_robust), and evaluated categorical inclusions. behavior_robust provided the most stable segmentation resistant to extreme outliers.

🧠 Model Selection

The project utilized a rigorous K-Means clustering tournament evaluating K=2 through K=10.

• Initial Heuristic Bias: K=2 initially appeared strongest because stability metrics (like ARI) heavily favor binary splits (e.g., Active vs Inactive).
• The Audit: A secondary manual audit proved that K=2 collapsed highly distinct engaged users (Power vs Streak vs Weekend) into a single monolithic group, defeating the product goals.
• Final Selection: `behavior_robust` + KMeans + K=4. This combination achieved near-perfect stability (ARI ~0.9998) across seeds and extracted highly actionable, balanced business segments. We accept the slight drop in Silhouette score compared to K=2 as a necessary tradeoff for granular product usefulness.

🎨 Persona Visuals

(Visualizations generated deterministically from the final pipeline)

Click to view Persona Distribution

Click to view Behavioral Radar Chart

Click to view Feature Heatmap

👥 Persona Details

💤 Dormant Learners (37.1%)

• Observed behavior: Show very low recent activity and substantially higher days since last active. Median days active last 30 is 0.
• Key evidence: Extremely low days_active_last_30 and sessions_per_week.
• Potential product opportunity: Push notifications with one-tap, highly simplified "refresher" content could safely re-engage users without overwhelming them.

🔥 Streak Maintainers (27.3%)

• Observed behavior: Exhibit frequent activity and unusually long streaks while maintaining relatively short sessions.
• Key evidence: Massive standardized effect size on longest_streak_days.
• Potential product opportunity: 1-minute "rapid review" modules or micro-lessons to support consistent engagement on busy days.

🌤️ Weekend Warriors (26.2%)

• Observed behavior: Concentrate their activity into fewer, longer sessions, mostly on weekends.
• Key evidence: Exceptionally high share_sessions_weekend.
• Potential product opportunity: Weekly progress goals may fit this behavior far better than restrictive daily streaks.

⚡ Power Learners (9.4%)

• Observed behavior: Exhibit substantially higher learning volume, XP accumulation, and leaderboard participation.
• Key evidence: Top decile across all volume and intensity metrics (XP, lessons, courses active).
• Potential product opportunity: Elite leaderboard tiers or advanced course tracks to increase long-term retention for highly active users.

🧪 Product Strategy

(These are proposed experiments, not validated causal effects.)

🔄 Pipeline Diagram

graph TD
    A[Raw Dataset] --> B(EDA)
    B --> C(Preprocessing)
    C --> D[Feature Representations]
    D --> E(K-Means Tournament)
    E --> F(Model Selection Audit)
    F --> G[K=4 Validation]
    G --> H(Persona Construction)
    H --> I(Evidence Audit)
    I --> J[Product Experiments]

🚀 Reproducibility

The repository is built for 100% determinism.

1. Clone the repo:

      git clone https://github.com/sadiyasyed28/Learner-Personas.git
      cd Learner-Personas

2. Environment: Create a virtual environment and install requirements:

      pip install -r requirements.txt

3. Dataset: Place the raw production dataset at the required location (dataset is not tracked in git due to data privacy policies).
4. Run tests:

      pytest tests/

5. Run pipelines: You can safely execute the Python scripts in src/ to regenerate the identical models, reports, and PNG visualizations.

📁 Project Structure

Learner-Personas/
├── data/
│   ├── processed/      # Preprocessed feature representations
│   └── outputs/        # Generated clusters, plots, evidence profiles
├── reports/
│   ├── final/          # Complete audit reports
│   ├── modeling/       # Clustering evaluation metrics
│   ├── personas/       # Extracted personas & summaries
│   └── product/        # Experiment specifications
├── src/                # Core ML pipeline and generation scripts
├── tests/              # 21 Pytest validations
├── requirements.txt    # Project dependencies
└── README.md

⚠️ Limitations

• Observational Data: This dataset captures observation behavior; clustering does not prove causality.
• Behavioral, Not Psychological: Personas are purely behavioral segments. We do not make claims about underlying motivations, demographics, or intelligence.
• K=4 is a Business Choice: While stable, K=4 is an analytical choice balancing mathematical separation and product usefulness.
• Hypothesis Testing Required: The proposed product interventions require rigorous A/B experimentation to validate causal effects on retention.
• Population Constraints: The dataset is a point-in-time snapshot and may not generalize to future cohorts without pipeline retraining.

✨ Technical Highlights

• 100% reproducible preprocessing pipelines.
• Robust handling of extreme outliers via median/IQR scalers.
• Rigorous clustering tournament evaluating multiple feature permutations across K=2..10.
• Deep model selection audit combating heuristic stability bias.
• Strict evidence-based persona construction avoiding inferential leaps.
• Fully automated Pytest suite asserting mathematical logic and output integrity.
• Detailed A/B experiment spec translation for Product teams.