Dynamic Heterogeneous Graph Model Boosts Calibrated Mobile Install Attribution

Researchers have developed ADAPT ClickNet, an end-to-end dynamic heterogeneous graph learning framework designed to solve the persistent challenge of mobile install attribution under severe class imbalance and strict temporal constraints. In mobile performance advertising, identifying the exceedingly rare clicks that result in genuine application installations from billions of log entries requires models to operate where positive events constitute a tiny fraction of total activity. Evaluating the system on an internal chronological test split of 25,734,126 click events from a large-scale public corpus, the researchers demonstrated that ADAPT ClickNet achieved an area under the receiver operating characteristic curve of 0.9856 and an area under the precision-recall curve of 0.7012. Furthermore, the model recorded an F1 score of 0.5328, a log loss of 0.00712, a precision at the top one percent of 0.2114, and a recall at the top one percent of 0.9311. These results reflect a substantial capacity to isolate rare attributed clicks without relying on future information.

The operational challenge of mobile install attribution stems from the difficulty of discriminating genuine conversion signals from redundant, high-frequency, or potentially fraudulent click traffic. Conventional attribution systems often rely on gradient-boosted decision trees, deep tabular architectures, sequence models, or static graph neural networks. However, these traditional approaches commonly treat click events as isolated static observations or examine temporal sequences only within narrow behavioral windows. Moreover, static models fail to capture the complex, multi-entity relationships linking internet protocol addresses, applications, hardware devices, operating systems, advertising channels, and temporal intervals. A critical vulnerability in prior modeling efforts is temporal data leakage, in which features inadvertently incorporate information from future timestamps, future entity frequencies, post-click conversion indicators, or noncausal graph neighborhoods. When temporal leakage occurs during training or evaluation, it creates artificially inflated predictive metrics that degrade precipitously once deployed in real-time scoring environments where only past observations are available.

To establish an empirically verifiable benchmark, the investigation utilized the TalkingData AdTracking corpus comprising 184,903,890 labeled click records. Across the complete corpus, direct aggregation identified 456,846 attributed clicks alongside 184,447,044 non-attributed clicks, establishing an extreme positive class prevalence of 0.247072 percent, which corresponds to a positive-to-negative class ratio of approximately 1 to 403.74. Rather than applying randomized sampling splits that mix events across chronological boundaries, the experimental protocol enforced strict chronological partitioning using fixed universal coordinated time thresholds. The self-supervised pretraining phase utilized an 82,258,712-record window spanning from November 6, 2017, at 16:00 UTC to November 8, 2017, at 04:00 UTC, completely withholding attribution labels. Supervised fine-tuning used 131,886,953 records, validation was performed on 27,282,811 records, and internal testing evaluated the final 25,734,126 records containing 63,993 attributed clicks. This temporal division ensured that all model training strictly preceded evaluation periods.

Data preprocessing was engineered to prevent both direct target leakage and subtle delayed-label leakage while optimizing memory usage. Integer columns were downcast to compact unsigned representations, reducing data size from 7.02 gigabytes to 1.91 gigabytes. All click records were sorted globally by timestamp before any feature construction. Crucially, the attributed time field was eliminated from the predictive feature space because post-click attribution timestamps cannot exist prior to conversion in production environments. Descriptive delay analysis revealed that conversion delays ranged from zero to 86,110 seconds, with a median delay of 293 seconds. To prevent delayed-label leakage, researchers defined an empirical 24-hour horizon of 86,400 seconds; Bayesian-smoothed target encodings for IP, app, device, operating system, and channel were computed strictly from historical records whose label-availability time preceded the target click timestamp. Additional causal predictors included past-only rolling counts computed across one-minute to one-day windows, time-since-previous-click recency measurements, and past-only frequency encodings.

The structural backbone of ADAPT ClickNet is a dynamic heterogeneous click graph that models relational context while enforcing strict causal directionality. Each click event is instantiated as a node alongside typed entity nodes representing IP address, app, device, operating system, channel, and 15-minute time buckets. All heterogeneous relations are explicitly directed: past clicks write to entity nodes, while separate read relations transmit summarized historical entity states to the target click, which acts as a read-only sink. To prevent graph density explosion, the researchers restricted temporal click-to-click edges exclusively to events sharing an identical IP address occurring within a preceding 60-second window, capped at the 25 most recent predecessors. An unconstrained shared-entity rule would have generated an estimated 8.3 trillion temporal edges dominated by high-frequency device identifiers; the shared-IP rule yielded approximately 2.0 billion directed edges. Graph construction utilized four consecutive 15-minute snapshots covering a 60-minute historical window.

ADAPT ClickNet combines three complementary representations through an adaptive tri-branch encoding architecture. The tabular entity branch maps high-cardinality categorical identifiers into 32-dimensional trainable embeddings while projecting causal temporal counts, recency indicators, and frequency statistics through a 128-dimensional dense transformation. The temporal graph branch processes the causal snapshots using a three-layer Heterogeneous Graph Transformer featuring four attention heads, a hidden dimension of 128, and neighbor-sampling fan-outs of 5, 10, and 25. The session-sequence branch groups clicks into deterministic sessions based on composite IP, device, and operating system keys separated by inactivity gaps exceeding 30 minutes, processing up to 32 consecutive events with a two-layer causal Transformer. Each branch outputs a 128-dimensional representation. Rather than using fixed addition or scalar gating, a gating layer computes click-specific three-way softmax fusion weights, enabling the network to dynamically emphasize tabular, graph, or sequential evidence based on the specific context of each click.

Because attribution labels are exceedingly sparse, ADAPT ClickNet incorporates multi-objective self-supervised pretraining across unlabeled click streams prior to supervised task adaptation. Pretraining optimizes a combined loss across five distinct auxiliary tasks: masked entity reconstruction, which replaces 15 percent of entity tokens and reconstructs them; temporal order prediction, which classifies the chronological permutation of session triplets; neighborhood reconstruction, which optimizes causal edge prediction against five negative samples per positive edge; session contrastive learning, which applies InfoNCE loss across independently augmented session views; and click-burst perturbation consistency, which enforces representation stability under edge dropping, feature masking, and timestamp jitter. Executed over five epochs across 65,873,425 records, pretraining required 3.1 GPU hours on a single NVIDIA A100 processor, achieving a final weighted loss of 2.670. Component losses reached 1.213 for masked reconstruction, 0.184 for order prediction, 0.962 for neighborhood reconstruction, 0.318 for session contrastive learning, and 0.170 for burst consistency.

Supervised fine-tuning adapts the pretrained representations using a multi-task loss function specifically designed to balance class disparity, ranking accuracy, and probability calibration. The loss function combines focal binary cross-entropy with a pairwise hinge ranking loss weighted at 0.3 and a differentiable calibration loss weighted at 0.1. The focal objective applies a focusing parameter of gamma equals 2.0 and alpha equals 0.75 to suppress gradients from easy negative samples and focus learning on hard instances. The pairwise ranking loss samples five negative clicks per positive click to enforce a margin of 1 between positive and negative logits. To ensure probability reliability, the differentiable calibration term computes soft memberships across 15 probability bins using Gaussian smoothing with temperature tau equals 0.05, minimizing the squared discrepancy between soft model confidence and observed empirical positive rates across mini-batches. Optimization utilized the AdamW optimizer with cosine learning-rate decay and bfloat16 mixed precision.

In comparative evaluations against 15 competitive baselines, ADAPT ClickNet demonstrated marked superiority over linear models, tree-based ensembles, deep tabular networks, sequence transformers, and graph architectures. LightGBM served as the strongest traditional tabular baseline, recording an area under the receiver operating characteristic curve of 0.9784, an area under the precision-recall curve of 0.6378, an F1 score of 0.4953, and a top one percent recall of 0.9007. ADAPT ClickNet surpassed LightGBM by 0.0634 in precision-recall AUC and 0.0304 in top one percent recall while lowering log loss from 0.00811 to 0.00712. Standalone Heterogeneous Graph Transformers and Sequence Transformers reached precision-recall AUCs of only 0.6285 and 0.6173, respectively, confirming that single modalities underperform integrated architectures. A controlled baseline augmenting LightGBM with causal graph descriptors achieved a precision-recall AUC of 0.6521, proving that access to graph features alone does not replicate the performance of end-to-end representation learning.

Reliability analyses demonstrated that ADAPT ClickNet achieved superior calibration and robust performance under varied operating conditions and simulated adversarial interference. The framework produced a Brier score of 0.001753 and an expected calibration error of 0.0064, outperforming LightGBM’s Brier score of 0.001883 and expected calibration error of 0.0103. Segment-wise evaluations across traffic intensity hours, device popularity tiers, and application conversion groups demonstrated consistent precision-recall AUC gains across all examined subsets. Under a mixed adversarial burst-injection stress test introducing synthetic click perturbations at a one percent rate, ADAPT ClickNet retained 93.18 percent of its clean precision-recall AUC, outperforming all baseline models in resilience. The authors noted essential operational limitations, emphasizing that evaluations relied on a single historical 2017 dataset, click attribution reflects installation records rather than audited fraud classifications, and practical production deployments must account for higher training runtimes and inference latencies relative to lightweight tree models.

Subject of Research: Leakage-controlled dynamic heterogeneous graph learning for mobile click install attribution under extreme class imbalance.

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

Article References: Chy, K. S., Islam, M. A., Chowdhury, M. S. I., Chowdhury, M. N. I., Islam, M. A., Islam, A., & Keneni, R. (2026). ADAPT ClickNet a leakage controlled dynamic heterogeneous graph learning framework for fraud aware mobile install attribution. Discover Computing, 29(1), Article 790. https://doi.org/10.1007/s10791-026-10619-7

Image Credits: AI Generated

DOI: 10.1007/s10791-026-10619-7

Keywords: Mobile Install Attribution, ADAPT ClickNet, Dynamic Heterogeneous Graphs, Graph Neural Networks, Click Fraud Detection, Temporal Data Leakage, TalkingData AdTracking, Class Imbalance, Heterogeneous Graph Transformer, Self-Supervised Pretraining, Probability Calibration, Session Modeling, Precision-Recall AUC, Adversarial Robustness, Machine Learning in Advertising, Brier Score