🏹 We Are Who We Lock: Overwatch Hero Psychology & Archetype Analysis

A Data Science Project for INST414 – Fall 2025 Author: Joshua Kwan

---

📌 Project Overview

This project extends my module-level Medium post “We Are Who We Lock: A Similarity Study of Overwatch Heroes & Player Psychology” into a full multi-method data science analysis combining:

• Cosine similarity (network science)
• Dimensionality reduction (PCA)
• Unsupervised learning (K-Means clustering)
• Evaluation metrics (silhouette score)
• Feature scaling, visualization, and interpretability techniques

The goal of this project is to answer a central question:

Do Overwatch heroes naturally form psychological archetypes—and can these archetypes help players find heroes that match their natural playstyle?

This repository contains all source code, datasets, charts, and analysis used to build the final Medium article.

---

🎯 Motivation

Players often gravitate toward specific heroes—mobile duelists, precision snipers, utility supports, or tanky frontline anchors.

This project attempts to quantify playstyle archetypes using hero attribute data and modern data science methods. The insights help:

• New players discover heroes aligned with their instincts
• Intermediate players understand their strengths and expand their roster
• Analysts visualize the hero ecosystem in a psychological feature space

---

📁 Repository Structure

Current tree (root-level files):

.
├── LICENSE
├── overwatch_full_similarity_top5.csv
├── overwatch_hero_archetype_analysis_final.ipynb
├── overwatch_hero_archetypes.csv
├── overwatch_hero_psychology_full.csv
└── README.md   ← You are here

Note: Some sections reference folders like data/, notebooks/, and images/. In this repo, assets and the notebook currently live at the root. If you’d like, we can organize them into folders for cleanliness.

---

📊 Methods Used

🔹 1. Cosine Similarity

Used to construct an N×N psychological similarity matrix between all heroes (N = number of heroes included).

🔹 2. PCA (Principal Component Analysis)

Reduces 4D hero features → 2D “hero psychology map.”

🔹 3. K-Means Clustering

Finds latent archetypes independent of hero role:

• Skirmishers
• Tacticians
• Anchors
• Sharpshooters

Validated with silhouette scoring.

🔹 4. Visualizations

Includes:

• PCA hero map
• Cosine similarity heatmap
• Radar charts (archetype profiles)
• Parallel coordinate plots
• Role distribution across clusters

---

🧪 How to Run

1. Clone the repository

git clone https://github.com/Kwanjk/updated-overwatch-heroes-player-analysis.git
cd updated-overwatch-heroes-player-analysis

2. Install dependencies

You can manually install or use a requirements file.

python -m pip install --upgrade pip
pip install pandas numpy matplotlib seaborn scikit-learn jupyter

3. Open the analysis notebook

jupyter notebook overwatch_hero_archetype_analysis_final.ipynb

4. Run all cells

The notebook will:

• Load the dataset
• Scale and preprocess features
• Compute PCA and cosine similarity
• Perform clustering
• Output all figures and CSV files

---

📈 Key Outputs

✔ Archetype Assignments

File: overwatch_hero_archetypes.csv Maps: Hero → Role → Cluster → Archetype name

✔ Top-5 Similar Heroes per Hero

File: overwatch_full_similarity_top5.csv

✔ PCA 2D Hero Map

Shows how heroes cluster organically in playstyle space.

✔ Final Radar Charts

Visual fingerprints of each archetype.

---

🧹 Data Cleaning & Bugs to Avoid

This project included:

• Normalizing numeric features
• Adjusting HP scale for comparability
• Ensuring consistent column naming
• Re-checking newly added heroes (e.g., Juno) as patches release

Common pitfalls others may run into:

• Forgetting to scale HP → breaks PCA
• Missing a hero in one-hot encoding → misaligns clustering
• Typos in dictionary → breaks DataFrame creation
• Using unscaled data in K-Means → distort clusters

---

🤖 AI Assistance Statement

ChatGPT was used for:

• Debugging code
• Improving visualization aesthetics
• Structuring the Medium article
• Generating descriptive labels and archetype names
• Ensuring scalability of the pipeline

All outputs were carefully checked, corrected, and validated manually for accuracy.

---

📜 Final Medium Article

🔗 Medium Link: Insert Medium article URL here once published.

This article contains the full narrative, explanations, and charts used to communicate the results.

---

⚠️ Limitations & Ethics

• Hero stat ratings are semi-subjective
• Patch changes may alter hero profiles
• PCA reduces complexity at the cost of nuance
• K-Means assumes spherical clusters (may oversimplify)
• No user gameplay data analyzed (privacy preserved)

Players should use this as guidance, not strict categorization.

---

🏁 Conclusion

This project demonstrates that Overwatch heroes naturally separate into statistically meaningful playstyle archetypes, reflecting real psychological differences in how players approach the game.

These archetypes help players:

• Better understand their strengths
• Explore new heroes confidently
• Build wider, more intuitive hero pools
• Reflect on their personal playstyle identity

Ultimately: We really are who we lock.