SAGE is trained on roughly 10,000 Steam games merged from
SteamSpy (ownership & review aggregates) and the
Steam Web API (store-page metadata). The pipeline
below shows exactly how the raw sources become the
steam_10k_prelaunch.csv file used by the model.
Provides aggregate ownership buckets, review counts, playtime
statistics, price, genre flags, and release month. This forms the
base table keyed by appid.
Per-game store-page metadata: descriptions, screenshots, trailers,
supported languages, platforms, categories, tags, packages, and
achievements. Loaded as a dict keyed by appid.
Two scripts produce the final training file. The first enriches and merges; the second selects pre-launch-only features and builds the classification target.
Read SteamSpy CSV into a DataFrame; load Steam Web API JSON
keyed by appid string.
For each game, derive ~30 columns: platform support, language counts, screenshot & trailer counts, description lengths, publisher/developer counts, Steam category flags (achievements, workshop, cloud, controller, VR…), tag counts and top-tag vote totals, package & SKU counts.
appid (left join)Base CSV ⟕ JSON features. Rows with no JSON match keep
their base columns; new numeric features are filled with
0.
Seven derived features combine raw signals into
interpretable proxies: store_page_score,
platform_reach, marketing_score,
publisher_backing, localization_score,
steam_integration, and an
is_mature_content flag.
Reviews, playtime, CCU, and owners themselves are excluded from the model — they wouldn't exist before launch and would leak the target.
Owners bucket → ordinal class 0–5 (top sparse buckets
merged into ≥750K). Median imputation for any
remaining numeric NaNs, then a stratified 80/20 train-test
split followed by z-score scaling.
The model predicts an ordinal owner tier, not a raw owner count. The original 9 SteamSpy buckets are collapsed to 6 because the top 4 had too few samples (n = 136, 88, 53, 19) for reliable learning.
| Class | Owner tier | Notes |
|---|---|---|
0 | ≤ 10K | Majority class (~69%) |
1 | ≤ 35K | |
2 | ≤ 75K | |
3 | ≤ 150K | |
4 | ≤ 350K | |
5 | ≥ 750K | Merged top 4 sparse buckets |
Only signals that exist before a game ships are used. Grouped by origin:
The full pipeline lives in two Python scripts. Output is the file the model trains on:
enrich_prelaunch.py
Merges SteamSpy CSV with Steam API JSON and engineers composite features.
train_prelaunch_model.py
Selects pre-launch features, builds the ordinal target, trains the stacked ensemble.
steam_10k_prelaunch.csv
Final 10,000 × 71 training file consumed by the model.