An improved graph factorization machine based on solving unbalanced game perception.

Area of Science:

• Data Science
• Machine Learning
• Human-Computer Interaction

Background:

• User perception is vital for mobile game user experience and profitability.
• Sparse game data and imbalanced categories challenge predictive model performance.
• Existing models struggle with high-dimensional features and data heterogeneity.

Purpose of the Study:

• To propose a novel method, the balanced graph factorization machine (BGFM), to address data imbalance and high-dimensional features in mobile game perception modeling.
• To enhance user experience prediction accuracy and, consequently, game profitability.

Main Methods:

• Data balancing using Borderline-SMOTE oversampling to address category imbalance.
• Graph-structured feature representation to naturally capture feature relationships.
• Integration of Factorization Machines (FM) and Graph Neural Networks (GNN) with an attention mechanism for sequential feature interaction analysis.

Main Results:

• The proposed BGFM method demonstrated superior performance compared to eight state-of-the-art models.
• Significant improvements were observed across key metrics: Area Under the Curve (AUC), precision, recall, and F-measure.
• BGFM effectively aggregates beneficial features through graph-based interaction mechanisms.

Conclusions:

• The balanced graph factorization machine (BGFM) offers a robust solution for modeling user perception in mobile games, particularly with sparse and imbalanced data.
• The hybrid FM-GNN approach with attention mechanism enhances the ability to capture complex feature interactions.
• BGFM significantly outperforms existing methods, paving the way for improved user experience and game monetization strategies.