Machine Learning the Steam Video Games Database

The Steam platform hosts over 27,000 games spanning dozens of genres. This thesis explores how those genres co-evolve, which are most influential, and what game properties best explain player engagement, using a mix of network science and machine learning.

The core contribution is a bipartite genre network connecting games to their genres, projected into a weighted, unipartite genre-to-genre graph. That graph is then analysed with centrality measures, Louvain community detection, and temporal slicing across five eras (1995–2020). A regression pipeline on game properties attempts to predict average playtime.

The pipeline had four sequential stages:

• EDA & Pre-processing: correlation matrix, distribution checks, genre explosion into separate rows.
• Bipartite Network: games and genres as two node types; edges represent game-genre membership. Projected to a unipartite genre network via matrix multiplication in Gephi's Multimode plugin.
• Network Analysis: centrality measures (degree, betweenness, closeness, eigenvector), temporal genre evolution across 5-year windows, Louvain community detection, node attributes (net ratings, median playtime).
• Regression: Ridge, Decision Tree, and Random Forest models trained to predict Average_Playtime. Features include dummy-encoded platforms, genres, categories, owners, price, achievements, and a computed ratings_ratio. 10-fold cross-validation for model selection.

Indie and Casual genres are connected to every other genre in the network (degree centrality = 1.0), meaning no genre on Steam exists in isolation from them.

Louvain modularity clustering at resolution 0.5 identified 8 communities:

• "Action" and "Indie" are the dominant parent genres; most emerging sub-genres trace back to one of them.
• Violent, Casual, Adventure, RPG, and Strategy all emerged from Action/Indie in the 2000s.
• "Free to Play" was first incorporated by Indie games, spawning the Massively Multiplayer cluster.
• Non-gaming genres (Education, Utilities, etc.) emerged independently but were later absorbed into mainstream gaming genres.
• "Early Access" appeared from 2010 onward as developers sought community feedback loops.
• Racing gave birth to the Sports sub-genre, a separate evolutionary branch from the Action/Indie lineage.

• Action games received the highest net ratings on the platform; Indie games came second.
• Indie games had the highest median playtime, likely due to their deep storylines and experimental gameplay that reward extended investment.
• The combination of high ratings and high playtime makes Indie + Action the twin pillars of Steam engagement.

Predicting Average_Playtime from game properties proved difficult. Most variability in engagement is explained by factors outside the available feature set. Ridge Regression was the best model by cross-validated RMSE.

10-fold cross-validation confirmed Ridge as the strongest generaliser:

The top feature for predicting average playtime was ratings_ratio (importance 0.20), followed by release year, achievements count, and price.

Honestly, the R² of ~2% surprised me at first. I expected static game properties to explain a decent chunk of playtime. But once I thought about it, it made sense: what actually drives engagement (streamer hype, patch updates, social trends, seasonal sales) is dynamic and community-driven. The dataset only captured snapshots of fixed properties. This was the biggest lesson from the project: sometimes the most valuable finding is learning that your features don't capture what matters, because it points you toward what does.

• Indie and Casual are the most structurally central genres; every other genre co-occurs with them.
• "Massively Multiplayer" and "Free to Play" have become synonymous, sharing player interaction models and monetisation strategies.
• Ratings are the strongest predictor of playtime; lower-priced or free-to-play games show higher engagement on average.
• Incorporating achievements actively boosts playtime by improving in-game motivation.
• Regression R² values stayed below 2%, suggesting community-driven factors (streamers, social trends, patches) dominate over static game properties.
• Class imbalance (few blockbusters vs. many niche titles) is the primary barrier to regression accuracy.

• The dataset cuts off at 2019, so the post-pandemic gaming boom and live-service genre growth are not captured.
• Adding game tag data alongside genre data could reveal finer sub-genre dynamics.
• Temporal regression (time-series of playtime per game) could outperform static predictors.
• Graph neural networks on the genre graph could learn richer genre interaction embeddings than centrality scores alone.