Temporal Graph Learning · Fraud Detection

Temporal GNN for Blockchain Fraud Detection

A temporal-graph research build comparing TGAT-style structural reasoning with an optimized XGBoost baseline under a documented benchmark protocol.

Documented benchmark · modelling trade-offs clearly labelled

Temporal-graph modelling trade-off

TGAT improves predictive quality in the documented benchmark run while requiring greater latency and memory than XGBoost.

Scope

Role and problem

My role: Built and documented a temporal-graph modelling pipeline and comparative validation path. Public claims are limited to the published benchmark artifacts.

Fraud and anomalous interactions are relational and time-dependent. Flattened tabular features can be fast, but they may lose higher-order structural context as interactions evolve across a network.

Architecture

System flow

Temporal interaction stream

Rolling tabular features

XGBoost baseline

Continuous-time encoding

TGAT-style attention

Dynamic link prediction

Latency and memory trade-off analysis

Results and scope

Public record

98.42%

Wikipedia-profile TGAT AUC-ROC

Benchmark result from the comparative report.

94.20%

Enterprise-profile TGAT AUC-ROC

Benchmark result on the higher-complexity transaction profile.

31.5–78.9 ms

TGAT latency range

Per-batch inference latency across the documented benchmark profiles; XGBoost remains substantially faster.

Signal	Scope	Interpretation
98.42%	Wikipedia-profile TGAT AUC-ROC	Benchmark result from the comparative report.
94.20%	Enterprise-profile TGAT AUC-ROC	Benchmark result on the higher-complexity transaction profile.
31.5–78.9 ms	TGAT latency range	Per-batch inference latency across the documented benchmark profiles; XGBoost remains substantially faster.

Shared here: Shared here: benchmark charts for modelling trade-offs and pipeline behaviour.

Published artifacts

Selected artifacts.

Charts, screenshots, and media artifacts supporting this case study.

Benchmark table comparing TGAT and XGBoost

image artifact

TGAT versus XGBoost benchmark table

Documented benchmark across Wikipedia interaction, Reddit hyperlink, and enterprise-style transaction profiles.

Bar chart comparing TGAT and XGBoost AUC ROC values

image artifact

TGAT and XGBoost AUC-ROC comparison

AUC-ROC comparison from the benchmark report.

Bar chart comparing TGAT and XGBoost inference latency

image artifact

Inference-latency trade-off

Per-batch latency comparison from the benchmark report. Lower is faster.

Contribution

Modelled dynamic interactions as a temporal graph.
Built a TGAT-style validation path with continuous-time encodings and attention-based message passing.
Compared predictive quality, latency, and memory trade-offs against a tabular XGBoost baseline.
Label the benchmark as a documented comparison, not a production deployment.

Lessons

Model choice should follow the structure of the problem.
Higher predictive quality can carry meaningful latency and memory costs.
Benchmark evidence is useful only when its boundary is explicit and the protocol is documented.

Limitations

The public benchmark is a documented comparison rather than a live production benchmark.
The enterprise transaction profile is a representative validation scenario, not a disclosed client dataset.
Deployment decisions still require dataset-specific reruns, calibration, and operational testing.

Stack

PyTorch
NetworkX
TGAT
Temporal GNNs
XGBoost
Fraud Detection
Benchmarking