🎮 PUBG Player Frustration Analysis

Predicting Match Placement and Identifying High-Value Players at Risk of Churn

📋 Table of Contents

• Overview
• Problem Statement
• Key Findings
• Dataset
• Methodology
• Technical Highlights
• Results
• Business Impact
• Installation
• Usage
• Project Structure
• Future Work
• Acknowledgments
• License
• Contact

---

🎯 Overview

This project analyzes over 4.4 million PlayerUnknown's Battlegrounds (PUBG) match records to predict player placement and identify a critical business problem: high-performing players experiencing unexpected losses due to unmeasurable factors.

By building regression models that achieve 80.6% R², we identify the 19.4% unexplained variance and use this as a "frustration score" to detect players at risk of churn. The analysis reveals that frustrated players are not underperforming—they're overperforming on all measurable metrics (63% more walkDistance, 66% more weapons, 557% more aggression) but still lose to factors outside their control.

Key Insight: The most frustrated 1% of players represent the game's highest-value segment—experienced, aggressive, engaged players who are statistically more likely to spend money and influence community sentiment.

---

💡 Problem Statement

In battle royale games, player experience is defined by more than just winning. A player who performs well but loses due to bad luck has a vastly different experience than a casual player who underperforms as expected.

Research Question:
Can we use in-match player actions to predict expected placement, then use prediction errors to quantify match frustration and identify players at churn risk?

Business Value:
Identifying frustrated high-performers allows game developers to:

• Implement targeted retention strategies
• Fine-tune game balance to reduce RNG impact
• Provide transparency through post-match analytics
• Reward performance beyond just placement

---

🔍 Key Findings

Model Performance

Model	MAE	RMSE	R²	Key Insight
Linear Regression	0.1213	0.1561	0.7416	Baseline; struggles with bounded predictions
Random Forest	0.0992	0.1362	0.8035	Handles non-linearity; reveals 92% feature importance in walkDistance
LightGBM	0.0986	0.1354	0.8056	Best performance with engineered features

Frustration Analysis

• 19.4% unexplained variance represents factors beyond player control (circle RNG, opponent skill, positioning)
• Frustrated players (bottom 1% of residuals) show paradoxical superiority:
  • +63% walkDistance (survival time)
  • +66% weapons acquired (looting effectiveness)
  • +557% aggression score (combat engagement)
• Yet they place 17 percentile points below predictions

Feature Importance

1. walkDistance (42.95%) - Survival time is king
2. walk_squared (34.21%) - Exponential survival advantage
3. walk_x_weapons (14.10%) - Looting effectiveness given survival time
4. Other features (<10%) - Combat, efficiency, playstyle modifiers

Strategy: Survive → Prepare → Win

---

📊 Dataset

Source: PUBG Finish Placement Prediction (Kaggle)

Size: 4,446,966 player records across ~47,000 matches

Target Variable:

• winPlacePerc: Continuous [0.0, 1.0] representing placement percentile

Key Features (29 total):

• Combat: kills, damageDealt, killStreaks, longestKill
• Survival: walkDistance, rideDistance, swimDistance
• Resources: weaponsAcquired, boosts, heals
• Team: assists, revives, DBNOs
• Meta: matchType (solo/duo/squad, FPP/TPP)

Data Quality:

• Only 1 missing value (winPlacePerc) in 4.4M rows
• 212 records (0.005%) removed for suspected cheating (>40 kills or >1.5 kills/min)
• killPlace removed due to data leakage (post-match ranking)

---

🔬 Methodology

1. Exploratory Data Analysis

• Target distribution analysis (bimodal: early deaths vs winners)
• Multicollinearity detection via VIF (removed assists, headshotKills, DBNOs)
• Outlier investigation (cheater detection, kills per minute analysis)
• Match-type stratification (solo vs duo vs squad correlations)

2. Feature Engineering

Interaction Features (conditional effectiveness):

walk_x_weapons = walkDistance × weaponsAcquired    # Looting efficiency
walk_x_kills = walkDistance × kills                # Combat effectiveness
walk_x_boosts = walkDistance × boosts              # Resource management

Efficiency Ratios:

kills_per_distance = kills / (walkDistance + 1)    # Combat efficiency
weapons_per_distance = weapons / (walkDistance + 1) # Looting efficiency

Playstyle Indicators:

aggression_score = (kills × 10 + damage/100) / (walkDistance + 1)
passive_score = walkDistance / (kills + 1)

Non-linear Transformations:

walk_squared = walkDistance²                        # Exponential survival advantage
log1p(rideDistance), log1p(longestKill)            # Diminishing returns

3. Preprocessing Pipeline

1. Remove missing values and outliers
2. Log-transform skewed features (rideDistance, longestKill)
3. Target encode matchType (mean winPlacePerc per category)
4. StandardScaler for all numerical features
5. Train/validation split (80/20)

4. Model Selection

• Experiment 1: Linear Regression (baseline)
• Experiment 2: Random Forest Regressor (handle non-linearity, multicollinearity)
• Experiment 3: LightGBM (efficiency + engineered features)

5. Frustration Score Calculation

frustration_score = actual_placement - predicted_placement
frustrated_players = bottom 1% of residuals (highest underperformance)

---

💻 Technical Highlights

Advanced Techniques Used

• ✅ Variance Inflation Factor (VIF) for multicollinearity detection
• ✅ Target Encoding for high-cardinality categorical variables
• ✅ Feature Interaction Engineering for conditional effectiveness
• ✅ Log Transformations for handling skewed distributions
• ✅ Residual Analysis for business insight generation
• ✅ Huber Loss for robust evaluation with outliers

Code Quality

• Modular preprocessing pipeline
• Comprehensive data validation
• Reproducible results (random_state=42)
• Detailed documentation and comments
• Visualizations for all key analyses

---

📈 Results

Feature Importance (LightGBM)

Feature	Importance
walkDistance	42.95%
walk_squared	34.21%
walk_x_weapons	14.10%
walk_x_boosts	2.54%
Others	<7%

Frustrated Player Profile

Average Player:           Frustrated Player:
• walkDistance: 1,154m    • walkDistance: 1,884m (+63%)
• weapons: 3.66           • weapons: 5.17 (+41%)
• kills: 0.92             • kills: 1.17 (+27%)
• aggression: 0.047       • aggression: 0.307 (+557%)

Expected placement: 52%   Actual placement: 35%
→ Disappointment gap: 17 percentile points

---

💼 Business Impact

Recommendations for Game Developers

A. Reduce RNG Impact (Make Skill Matter More)

1. Predictable Circle Mechanics

   • Show next 2 zones instead of 1
   • Weight circle spawns toward current zone center (70% probability)
   • Add probability heatmaps for next zone

2. Skill-Based Matchmaking

   • Implement hidden MMR system
   • Reduce "ran into a pro" scenarios
   • A/B test loose SBMM (±200 MMR range)

3. Consistent Loot Distribution

   • Guarantee high-tier loot zones
   • Dynamic loot balancing (if 5 buildings = no scope, next guarantees one)

B. Transparency & Feedback

1. Post-Match Analytics Dashboard

   📊 YOUR STATS:
   Placement: 35th percentile
   Walk Distance: 1,884m (Top 25% ⭐)
   Weapons: 5 (Top 40% ⭐)
   
   🎯 PREDICTED: 52nd percentile
   You SHOULD have placed better!
   
   ❌ WHAT WENT WRONG:
   • Circle spawned away from you 3 times (bad luck)
   • Encountered top 5% skilled opponent
   • Final circle positioning: 8th percentile
   
   💡 YOU PLAYED WELL! Focus on positioning.
   

2. Positive Loss Messaging

   • Replace "Rank 65/100" with "Top 25% survival, tough luck this round!"

C. Reward Performance Beyond Placement

1. Multi-Dimensional Ranking

   • Placement MMR (current)
   • Combat MMR (kills, damage)
   • Survival MMR (walkDistance, efficiency)

2. Performance-Based Bonuses

   • Top 20% walkDistance: +50 BP
   • Top 20% combat: +50 BP
   • "Overperformed prediction": +100 BP

Expected Impact

• 15-30% reduction in churn among high-performing players
• Protect 2-5% of revenue (frustrated players are high-engagement = high-LTV)
• Reduce brand risk from influencer/streamer frustration
• Improve perceived fairness across all player segments

---

🛠️ Installation

Prerequisites

• Python 3.8+
• Jupyter Notebook
• 8GB+ RAM (for full dataset)

Setup

# Clone repository
git clone https://github.com/dganesh05/pubg-frustration-analysis.git
cd pubg-frustration-analysis

# Create virtual environment
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

# Download dataset
# Visit: https://www.kaggle.com/c/pubg-finish-placement-prediction
# Download train_V2.csv and place in project root

Dependencies

pandas>=1.3.0
numpy>=1.21.0
scikit-learn>=1.0.0
lightgbm>=3.3.0
matplotlib>=3.4.0
seaborn>=0.11.0
category-encoders>=2.3.0
scipy>=1.7.0
statsmodels>=0.13.0
jupyter>=1.0.0

---

🚀 Usage

Quick Start

jupyter notebook pubg_analysis.ipynb

Analyzing Your Own Data

# Load your PUBG match data
import pandas as pd
df = pd.read_csv('your_data.csv')

# Run preprocessing
from preprocessing import clean_data, engineer_features
df_clean = clean_data(df)
df_eng = engineer_features(df_clean)

# Train model
from models import train_lgb_model
model = train_lgb_model(df_eng)

# Calculate frustration scores
from analysis import calculate_frustration
frustrated = calculate_frustration(model, df_eng)

Reproducing Results

All analyses are fully reproducible with random_state=42 set throughout the notebook. You can set a different random state by changing the RANDOM_STATE variable

---

📁 Project Structure

pubg-frustration-analysis/
│
├── README.md                 # This file
├── requirements.txt          # Python dependencies
├── project_notebook.ipynb       # Main analysis notebook
├── requirements.txt        # minimum requirements to download
└── train_V2.csv         # Dataset (download separately)

---

🔮 Future Work

Model Improvements

• Collect positioning data (distance to circle center, final zone position)
• Add opponent skill features (average opponent MMR in match)
• Implement temporal analysis (frustration cumulation over sessions)
• Test ensemble methods (stacking RF + LightGBM + XGBoost)

Business Applications

• A/B test recommendations (circle mechanics, SBMM, rewards)
• Build real-time frustration detection system
• Create player retention dashboard for developers
• Extend analysis to other battle royale games

Technical Enhancements

• Deploy model as REST API
• Build interactive web dashboard (Streamlit/Dash)
• Optimize for larger datasets (Dask, distributed computing)
• Add automated retraining pipeline

---

🙏 Acknowledgments

• Dataset: PUBG Corporation via Kaggle Competition
• AI Assistance: Claude (Anthropic AI) for feature engineering ideation and code debugging
• Libraries: scikit-learn, LightGBM, pandas, matplotlib, seaborn
• Inspiration: Data-driven game design and player psychology research

---

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

Note: The PUBG dataset is subject to Kaggle's competition rules and PUBG Corporation's terms of use.

---

📧 Contact

Divya Ganesh
📧 [email protected] 🔗 LinkedIn
💼 GitHub

---

🌟 If You Found This Useful

• ⭐ Star this repository
• 🐛 Report issues
• 🤝 Submit pull requests
• 💬 Share feedback

---

Built with ❤️ for data-driven game design