Hugging Face's logo

Datasets:
lukeslp
/
steam-network-data

Tasks:
Graph Machine Learning
Feature Extraction
Languages:
English
Size:
10K<n<100K
Tags:
steam
games
network-analysis
graph-data
collaborative-filtering
co-review
License:
Dataset card Files
xet
 Community
steam-network-data / README.md
lukeslp's picture
lukeslp
Upload README.md with huggingface_hub
e5e4bfb verified
preview code
|
raw
history blame contribute delete
3.82 kB
metadata 
license: cc-by-4.0
task_categories:
  - graph-ml
  - feature-extraction
language:
  - en
tags:
  - steam
  - games
  - network-analysis
  - graph-data
  - collaborative-filtering
  - co-review
  - video-games
pretty_name: Steam Universe Co-Review Network
size_categories:
  - 10K<n<100K
source_datasets:
  - FronkonGames/steam-games-dataset

Steam Universe Co-Review Network

A graph dataset connecting 25,996 Steam games through 861,232 weighted edges based on shared player reviews. If two games have many of the same reviewers, they're linked. The edge weight is the number of shared reviewers.

Paired with a full catalog of 82,928 Steam games (2005-2025) with genres, tags, ratings, prices, and developer info.

Live visualization: dr.eamer.dev/datavis/interactive/steam-network

Files

steam_network.json (41 MB)

The co-review graph.

  • 25,996 nodes (games with enough reviews to form connections)
  • 861,232 edges (weighted by shared reviewer count)

Node fields:

{
  "id": "1000010",
  "title": "Crown Trick",
  "year": "2020",
  "rating": "Very Positive",
  "ratio": 85,
  "reviews": 5263,
  "price": 4.99,
  "genres": [3, 0, 6, 5],
  "tags": [40, 7],
  "developer": "NEXT Studios"
}

Link fields: source (node index), target (node index), weight (shared reviewers).

steam_all_2005.json (6.2 MB)

Full catalog of 82,928 games in packed array format for compact transfer:

[name, year, approval_ratio, review_count, price, rating_index, genre_indices, tag_indices, developer]

Includes lookup tables for 9 rating tiers, 33 genres, and 50 tags.

steam_force_layout.json (252 KB)

Pre-computed force-directed layout coordinates. Saves about 30 seconds of simulation time when loading the visualization.

Pipeline scripts

Four Python scripts to rebuild the dataset from source:

Script Purpose
build_network_v2.py Scans 2GB recommendations.csv to find co-reviewers and build the edge list
enrich_data.py Processes the FronkonGames enriched CSV into the compact JSON catalog
compute_layout.py Runs a force simulation in Python to pre-compute node positions
build_all_games.py Legacy catalog builder (superseded by enrich_data.py)

Network Construction

The network is built from 41 million Steam user review records. Two games share an edge when 5 or more users reviewed both. Key parameters:

  • MIN_SHARED = 5 -- minimum shared reviewers for an edge
  • TOP_K = 50 -- maximum neighbors retained per node
  • MAX_USER_GAMES = 75 -- caps per-user pair generation to prevent combinatorial blowup

Use Cases

  • Game recommendations -- collaborative filtering through graph traversal
  • Community detection -- find genre clusters, indie vs. AAA ecosystems
  • Network analysis -- centrality measures, bridge games connecting disparate genres
  • Market analysis -- price, rating, and genre distributions across 82K titles
  • Visualization -- the companion interactive viz has 8 rendering modes

Sources

  • Game metadata: FronkonGames/steam-games-dataset (January 2026 snapshot)
  • User reviews: Kaggle Steam recommendations.csv (41M review records)
  • Network: Derived from the review data

Quick Stats

Metric Value
Games in catalog 82,928
Network nodes 25,996
Network edges 861,232
Genres 33
Tags 50
Rating tiers 9
Year range 2005-2025
Most reviewed Counter-Strike 2 (8.8M reviews)

Author

Luke Steuber