Table 5 Ablation study results across models and Datasets. Initially we tried to imnplemnt on Telco dataset, but this dataset does not perform well on graph based model, so we tried with KKBOX which works extensively for kkbox.

From: Advanced customer churn prediction for a music streaming digital marketing service using attention graph-based deep learning approach

Model

Dataset

Techniques

Test accuracy

Test AUC

Baseline GNN

Telco

Graph Sage, synthetic graph nodes

74.0%

0.82

Hybrid GNN

Telco

Feature engineering, SAGE, hybrid (Concat)

75.5%

0.83

HybridGAT (proposed)

KKBox

GAT, weighted loss, heavy regularization

88.4%

0.90

Ensemble (GAT + XGBoost)

KKBox

GAT embeddings as features for XGBoost

89%

0.90

HybridGAT (final model)

KKBox

Large-scale data, feature Eng., GAT, weighted loss

95.8%

0.963

Back to article page