KDD

\documentclass[sigconf,screen]{acmart}
\usepackage{graphicx}
\usepackage{amsmath,amssymb}
\usepackage{booktabs}

\acmConference[KDD '26]{ACM SIGKDD Conference on Knowledge Discovery and Data Mining}{August 2026}{City, Country}
\acmISBN{978-1-4503-XXXX-X/26/08}
\acmDOI{10.1145/XXXXXXX.XXXXXXX}
\setcopyright{acmlicensed}
\copyrightyear{2026}
\acmYear{2026}

\begin{document}

\title{Temporal Graph Representation Learning\\for Fraud Detection at Scale}

\author{First Last}
\affiliation{\institution{University of Example}\country{Country}}
\email{[email protected]}
\author{Jane Doe}
\affiliation{\institution{Example Research Labs}\country{Country}}
\email{[email protected]}
\author{John Smith}
\affiliation{\institution{University of Example}\country{Country}}
\email{[email protected]}
\renewcommand{\shortauthors}{Last et al.}

\begin{abstract}
Fraud detection in payment networks is a graph problem whose topology
evolves minute-by-minute. We present a temporal graph neural network
trained on 4B-edge billing graphs that improves recall at 1\% false
positive rate by 18\% over strong industrial baselines, while halving
inference latency through a hierarchical sampling scheme.
\end{abstract}

\begin{CCSXML}
<ccs2012><concept><concept_id>10010147.10010257</concept_id>
<concept_desc>Computing methodologies~Machine learning</concept_desc>
<concept_significance>500</concept_significance></concept></ccs2012>
\end{CCSXML}
\ccsdesc[500]{Computing methodologies~Machine learning}
\keywords{graph neural networks, fraud detection, temporal graphs}
\maketitle

\section{Introduction}
Payment fraud adapts to defenses in minutes. Static graph methods lag;
recurrent methods scale poorly; temporal GNNs sit in an awkward middle.

\section{Related Work}
TGN, JODIE, DyRep, industrial fraud systems.

\section{Method}
We construct a per-minute time-sliced graph and run a cross-slice
attention module over sampled neighborhoods.

\subsection{Hierarchical Sampling}
We bias sampling toward recently active neighborhoods and long-lived
high-risk patterns.

\section{Experiments}
\begin{table}[t]
\centering
\begin{tabular}{lccc}
\toprule
Method & Recall@1\%FPR & p50 latency (ms) & p99 \\
\midrule
Rules         & 0.51 &  3 &  8 \\
GraphSAGE     & 0.64 &  8 & 26 \\
TGN           & 0.71 & 22 & 68 \\
\textbf{Ours} & \textbf{0.84} & \textbf{11} & \textbf{34} \\
\bottomrule
\end{tabular}
\caption{Offline evaluation on internal payments data.}
\end{table}

\section{Deployment}
We describe A/B test results from a deployment covering \$300B annual GMV
across three business units.

\section{Conclusion}
Temporal GNNs can be deployed at industrial scale with careful attention
to sampling, batching, and latency guarantees.

\bibliographystyle{ACM-Reference-Format}
\bibliography{refs}
\end{document}