ADAPT ClickNet a leakage controlled dynamic heterogeneous graph learning framework for fraud aware mobile install attribution

Mobile install attribution requires identifying extremely rare attributed clicks from massive click logs under severe class imbalance, temporal drift, and leakage risk. Existing tabular, transformer, graph, and sequence-based click-prediction models commonly treat click records as static observations, exploit limited temporal interaction structure, weakly model heterogeneous relations among IP, app, device, OS, channel, time bucket, and click-event entities, and provide insufficient calibration for probability-based decision-making. This paper proposes ADAPT ClickNet, a leakage-controlled dynamic heterogeneous graph-learning framework for calibrated mobile install attribution. The framework integrates causal temporal feature construction, directed temporal click-graph construction, heterogeneous click–entity interaction modeling, a tri-branch encoder with tabular entity, temporal graph, and session-sequence branches, self-supervised click representation learning, click-specific three-way softmax fusion, and calibrated supervised fine-tuning. Self-supervised pretraining combines masked entity reconstruction, temporal order prediction, neighborhood reconstruction, session contrastive learning, and click-burst perturbation consistency. Supervised fine-tuning jointly optimizes focal binary cross-entropy, an AUC-oriented ranking loss, and a calibration loss to improve rare-event ranking and probability reliability. Evaluation is performed using chronological partitioning of a large-scale click log containing 184,903,890 labeled records. The internal test set contains 25,734,126 clicks, including 63,993 attributed clicks and 25,670,133 non-attributed clicks. ADAPT ClickNet achieves a ROC-AUC of 0.9856, PR-AUC of 0.7012, log loss of 0.00712, F1 score of 0.5328, precision at the top 1% of 0.2114, and recall at the top 1% of 0.9311, outperforming LightGBM, XGBoost, CatBoost, tabular transformers, graph neural networks, HGT, and sequence transformers. Ablation results show that leakage-controlled temporal features, self-supervised pretraining, graph encoding, tabular encoding, sequence encoding, adaptive fusion, ranking loss, and calibration loss each contribute to performance. Calibration analysis further yields a Brier score of 0.001753 and an expected calibration error of 0.0064. These findings support ADAPT ClickNet as an effective internally validated framework for leakage-controlled rare mobile install attribution with strong ranking performance and calibrated probability estimation.