Enhanced Semantic Relationship-Based Mobile App Description Classification Using Hybrid Graph Neural Networks and Advanced Data Science Techniques

Abstract

Due to the exponential growth in the number of mobile applications, correctly categorizing app descriptions into genres is essential for enhancing the user experience, discoverability, and personalization. Traditional text classification models struggle to represent the intricate semantic relationships within descriptions and often miss complex multi-genre issues and context- dependent nuances. To address these problems, we introduce a classification framework using Graph Neural Networks (GNNs) with improved semantic representation, supported by deep learning, data science, and reinforcement learning. Our approach begins with a Hybrid Graph Convolutional Network (GCN) paired with a Support Vector Machine (SVM): it captures semantic relationships among words as graph nodes to improve SVM-based classification margins. This hybrid method improves structural learning and yields accurate genre classifications, expected to perform 5-8% better than Graph Convolutional Networks (GCNs) in isolation. In the final integration, we combine a Contextual Graph Attention Network (CGAT) with Bi-directional Encoder Representations from Transformers (BERT) embeddings to capture rich, complex contextual relationships; we expect the model to achieve 92-94% accuracy for multi-genre descriptions. We improve data diversity through topic modeling with Non-negative Matrix Factorization and semantic data augmentation for thematic components to enhance the generalization and explainability of the models. Furthermore, SHapley Additive exPlanations (SHAP) explains model decisions by quantifying the contributions that words make to genre predictions, bringing much-eeded.