Lightweight_Transformer_Based_Sentiment_Analysis_of_Steam_Game_Reviews_Using_an_Ensemble_Approach
Description
This dataset contains 15,000 user-submitted reviews collected from the Steam platform via the official Steam Web API, spanning 30 distinct games, intended for binary sentiment classification research. Each review is paired with the reviewer's binary recommendation outcome (recommended / not recommended) as provided natively by Steam, along with associated review metadata. Following collection, the raw dataset underwent cleaning consisting of duplicate removal and invalid-review filtering, reducing the corpus from 15,000 to 11,400 usable reviews. User-identifying information was removed during preprocessing prior to release. The cleaned dataset was partitioned into training (7,980 reviews; 70%), validation (1,710 reviews; 15%), and test (1,710 reviews; 15%) splits using stratified sampling to preserve class balance across partitions. Train/validation/test partition integrity (absence of cross-split duplication) was independently verified via multiple methods (exact-match, TF-IDF cosine similarity, and Jaccard shingle overlap). The dataset exhibits class imbalance representative of naturally occurring Steam review distributions (approximately 86% positive, 14% negative across the full 11,400-review corpus).
Files
Steps to reproduce
A list of 30 Steam game titles was selected, with each game's unique Steam application ID stored in a reference file. For each of the 30 games, review data was collected via the official Steam Web API, requesting approximately 500 reviews per game, yielding a raw corpus of 15,000 reviews. The raw corpus was cleaned by removing duplicate reviews and invalid/empty entries, reducing the dataset to 11,400 reviews. User-identifying information was removed from all retained reviews during standardization, prior to any further processing. The cleaned, anonymized corpus (anonymous_reviews.csv) was partitioned into training (70%), validation (15%), and test (15%) splits using stratified sampling on the binary recommendation label, to preserve class balance across all three partitions. Partition integrity was verified to confirm no review text is duplicated across the training, validation, and test splits. Data collection, cleaning, standardization, splitting, and partition-integrity verification scripts are available at: https://github.com/Hayha-0/Lightweight_Transformer_Based_Sentiment_Analysis_of_Steam_Game_Reviews_Using_an_Ensemble_Approach