Ferdi Setyo Handika, Lili Dwi Yulianto, Septi Andryana

The Feature Selection vs Dimensionality Reduction for Steam Game Metadata Classification: An Ensemble Learning Study

Introduction

The feature selection vs dimensionality reduction for steam game metadata classification: an ensemble learning study. Optimize Steam game metadata classification. This study compares feature selection (MI) vs. dimensionality reduction (PCA, LDA) using ensemble learning. FS proves superior for predictive integrity, vital for game developers.

Abstract

Optimizing noisy Steam game metadata is essential for accurate binary classification. This study compares feature selection (MI) and dimensionality reduction (PCA, LDA) using a dataset of 55,144 Steam reviews and four ensemble algorithms, evaluated through Stratified 5-Fold Cross-Validation. The results show that the 125-feature baseline achieved the highest accuracy of 0.7728 with CatBoost. Feature selection (FS_10) maintained competitive performance with an accuracy of 0.7449, while LDA, after optimization, achieved 0.7281. In contrast, PCA significantly hindered performance (0.6963) due to the inability of linear transformations to preserve the discriminative power of one-hot encoded categorical features, which ensemble models handle better in their original form. These findings highlight the importance of preserving original features, especially in low-to-medium dimensional metadata, where feature selection outperforms dimensionality reduction in maintaining predictive integrity. High accuracy is crucial for developers to track product reception and for platforms to improve recommendation systems that influence user purchasing decisions. The study concludes that for Steam game metadata, strategic feature selection is superior to dimensionality reduction for maintaining classification performance.

Full Text

Comments

You need to be logged in to post a comment.