← Back to work

Temporal Graph Learning

Temporal GNN for Blockchain Fraud Detection

A graph-learning research build on modelling fraud as a relational and time-dependent problem.

Temporal graph-learning research build

t0t1t1t2t3t3fraud?

Conceptual temporal-graph illustration

The diagram explains the model choice without claiming an unpublished benchmark outcome.

Scope

Role and problem

My role: Built the temporal graph-learning pipeline and XGBoost comparison path.

Fraud is relational and time-dependent. Static tabular features can miss how transactions evolve across a network.

Architecture

System flow

01

Transaction records

02

Dynamic graph construction

03

Temporal encoding

04

Attention-based message passing

05

Dual-task learning

06

XGBoost comparison path

07

Evaluation rerun

Evidence

Measured signals

TGAT-style

Temporal architecture

Temporal encodings and attention-based message passing.

Dual-task

Learning design

Fraud-oriented temporal graph-learning pipeline.

XGBoost

Comparison path

A traditional benchmark route is included for reproducible comparison.

Public scope: The public case study documents architecture and benchmark design without claiming unpublished comparative performance.

Contribution

  • Modelled blockchain transactions as a dynamic graph.
  • Implemented temporal encodings and attention-based message passing.
  • Created a benchmark path against XGBoost without pretending an unpublished comparison is evidence.

Lessons

  • Model choice should follow the structure of the problem.
  • A benchmark is only valuable when the protocol is reproducible.
  • Architecture and benchmark claims should remain clearly separated.

Limitations

  • No comparative TGAT-versus-XGBoost performance claim is reported publicly.
  • The graph illustration is conceptual rather than dataset-specific.
  • The public scope is methodological.

Stack

  • PyTorch
  • NetworkX
  • TGAT
  • Temporal GNNs
  • XGBoost
  • Fraud Detection