lukeslp/steam-co-review-network

---
license: cc-by-4.0
task_categories:
- graph-ml
- feature-extraction
language:
- en
tags:
- steam
- games
- network-analysis
- co-review
- graph-data
- gaming
- social-network
pretty_name: Steam Co-Review Network
size_categories:
- 1M<n<10M
dataset_info:
features:
- name: nodes
dtype: list
- name: links
dtype: list
- name: meta
dtype: dict
---

# Steam Co-Review Network

Two Steam games share an edge when multiple users reviewed both. Built from 128 million user reviews across 80,000 games (2012-2024), this dataset maps how the Steam catalog is connected through player overlap.

## Files

### steam_network_full.json

The complete co-review graph with minimal filtering:

- 48,362 game nodes (every game with 10+ reviews)
- 33,041,298 weighted edges (2+ shared reviewers per pair)
- Per-node cap of 50 neighbors (densest connections preserved)
- Edge weights range from 2 to 420,410 shared reviewers

**Node format:**
```json
{
"id": "620",
"title": "Portal 2",
"year": "2011",
"rating": "Overwhelmingly Positive",
"ratio": 97,
"reviews": 263842,
"price": 9.99
}
```

**Link format:**
```json
{"source": 0, "target": 42, "weight": 1523}
```

Source and target are indices into the nodes array. Weight is the number of users who reviewed both games.

### steam_all_2005.json

82,928 games released 2005-2025. Packed as arrays for compact JSON:

```
[name, year, ratio, reviews, price, ratingIdx, genreIdxs, tagIdxs, developer]
[0] [1] [2] [3] [4] [5] [6] [7] [8]
```

Genres and tags are stored as index arrays referencing top-level `genres[]` and `tags[]` lookup tables in the same file.

### steam_force_layout.json

Genre-aware pre-computed layout positions for the top ~9K nodes, clustered by primary genre with hub games as anchors. Use as warm-start coordinates for force-directed visualization.

## Sources

- **Game metadata**: [FronkonGames Steam Games Dataset](https://huggingface.co/datasets/FronkonGames/steam-games-dataset) — Jan 2026 snapshot, 122K games
- **User reviews**: [artermiloff Steam Reviews 2024](https://www.kaggle.com/datasets/artermiloff/steam-games-reviews-2024) — 128M reviews across 80K games, one CSV per game, 2012-June 2024

## Pipeline

1. Load game metadata from FronkonGames enriched CSV
2. Scan 30K+ per-game review CSVs, extract steamid-to-game mappings
3. For each user who reviewed 2+ games, generate all game pairs
4. Count shared reviewers per pair to produce edge weights
5. Filter: minimum 2 shared reviewers (no neighbor cap)

Full pipeline: [github.com/lukeslp/steam-network-data](https://github.com/lukeslp/steam-network-data)

## Use Cases

- **Graph ML**: Node classification (predict genre/rating from network position), link prediction, community detection
- **Recommendation systems**: Games connected by high edge weights share audiences
- **Market analysis**: Which genres cluster together? Where are the gaps?
- **Visualization**: Force-directed layouts, chord diagrams, genre timelines of the Steam ecosystem

## Live Visualization

[dr.eamer.dev/datavis/interactive/steam/](https://dr.eamer.dev/datavis/interactive/steam/)

Four interactive Canvas-rendered views: universe scatter, chord diagram, force-directed network, and genre timeline.

## Distribution

- **GitHub**: [lukeslp/steam-network-data](https://github.com/lukeslp/steam-network-data)
- **Kaggle**: [lucassteuber/steam-universe-network](https://www.kaggle.com/datasets/lucassteuber/steam-universe-network)

## Author

Luke Steuber — [lukesteuber.com](https://lukesteuber.com) — [@lukesteuber.com on Bluesky](https://bsky.app/profile/lukesteuber.com)

license: 知识共享署名4.0许可协议（CC BY 4.0）
task_categories:
- 图机器学习（Graph ML）
- 特征提取（Feature Extraction）
language:
- 英语（English）
tags:
- Steam平台（Steam）
- 游戏（Games）
- 网络分析（Network Analysis）
- 协同评论（Co-Review）
- 图数据（Graph Data）
- 游戏行业（Gaming）
- 社交网络（Social Network）
pretty_name: Steam协同评论网络（Steam Co-Review Network）
size_categories:
- 100万<n<1000万
dataset_info:
features:
- name: 节点（nodes）
dtype: 列表（list）
- name: 边（links）
dtype: 列表（list）
- name: 元数据（meta）
dtype: 字典（dict）

# Steam协同评论网络（Steam Co-Review Network）

当多名用户同时对两款Steam游戏进行评论时，二者之间便会建立一条边。本数据集基于2012年至2024年间8万款游戏的1.28亿条用户评论构建，旨在通过玩家重叠情况映射Steam游戏商店的关联网络。

## 文件

### 完整协同评论网络文件：steam_network_full.json

该文件为经过最小化过滤的完整协同评论图（Co-Review Graph）：
- 共48362个游戏节点（包含至少10条评论的所有游戏）
- 33041298条带权边（每对游戏至少有2名共同评论用户）
- 每个节点最多保留50个邻居节点（保留最紧密的连接关系）
- 边权值范围为2至420410（代表共同评论用户的数量）

**节点格式：**
json
{
"id": "620", // 游戏的Steam应用ID
"title": "Portal 2", // 游戏名称
"year": "2011", // 发售年份
"rating": "Overwhelmingly Positive", // 玩家评价：压倒性好评
"ratio": 97, // 好评率百分比
"reviews": 263842, // 总评论数
"price": 9.99 // 售价（单位：美元）
}

**边格式：**
json
{"source": 0, "target": 42, "weight": 1523}

其中source与target为nodes数组中的索引值，weight代表同时评论两款游戏的用户数量。

### 全量游戏文件：steam_all_2005.json

包含2005年至2025年发售的82928款游戏，为实现JSON压缩采用数组打包格式：

[name, year, ratio, reviews, price, ratingIdx, genreIdxs, tagIdxs, developer]
[0] [1] [2] [3] [4] [5] [6] [7] [8]

翻译为：

[游戏名称, 发售年份, 好评率, 总评论数, 售价, 评价索引, 游戏类型索引, 标签索引, 开发商]
[0] [1] [2] [3] [4] [5] [6] [7] [8]

游戏类型与标签以索引数组形式存储，指向当前文件中顶层的`genres[]`与`tags[]`查找表。

### 预计算力导向布局文件：steam_force_layout.json

针对约9000个核心节点生成的、基于游戏类型的预计算布局坐标，节点按主力游戏类型聚类，并以头部游戏作为锚点。可作为力导向可视化的初始坐标使用。

## 数据源

- **游戏元数据**：[FronkonGames Steam游戏数据集（FronkonGames Steam Games Dataset）](https://huggingface.co/datasets/FronkonGames/steam-games-dataset) — 2026年1月快照，包含12.2万款游戏
- **用户评论数据**：[artermiloff 2024年Steam评论数据集（artermiloff Steam Reviews 2024）](https://www.kaggle.com/datasets/artermiloff/steam-games-reviews-2024) — 涵盖2012年至2024年6月期间8万款游戏的1.28亿条评论，每款游戏对应一个CSV文件

## 数据处理流程

1. 从FronkonGames扩充后的CSV文件中加载游戏元数据
2. 扫描3万余个单游戏评论CSV文件，提取Steam用户ID到游戏的映射关系
3. 针对每位评论了至少2款游戏的用户，生成其评论过的所有游戏配对
4. 统计每对游戏的共同评论用户数量，以此生成边权值
5. 过滤规则：仅保留至少有2名共同评论用户的游戏配对（未设置邻居节点上限）

完整处理流程：[github.com/lukeslp/steam-network-data](https://github.com/lukeslp/steam-network-data)

## 应用场景

- **图机器学习（Graph ML）**：节点分类（基于网络位置预测游戏类型与评价）、链路预测、社区检测
- **推荐系统**：边权值较高的游戏拥有重叠的玩家群体，可用于游戏推荐
- **市场分析**：分析哪些游戏类型存在聚类现象，以及市场空白点
- **可视化**：可用于Steam生态系统的力导向布局、弦图、游戏类型时间线可视化

## 在线可视化工具

[dr.eamer.dev/datavis/interactive/steam/](https://dr.eamer.dev/datavis/interactive/steam/)
该工具包含四种基于Canvas渲染的交互式视图：宇宙散点图、弦图、力导向网络图与游戏类型时间线。

## 分发渠道

- **GitHub**：[lukeslp/steam-network-data](https://github.com/lukeslp/steam-network-data)
- **Kaggle**：[lucassteuber Steam宇宙网络数据集（lucassteuber/steam-universe-network）](https://www.kaggle.com/datasets/lucassteuber/steam-universe-network)

## 作者

卢克·斯托伊伯（Luke Steuber） — [lukesteuber.com](https://lukesteuber.com) — [Bluesky平台@lukesteuber.com](https://bsky.app/profile/lukesteuber.com)