报告详情
Learned Meta-Fusion for Credit Card Fraud Triage: A Segment-Conditional Extension of Composite Risk Scoring
编号：32
访问权限：仅限参会人
更新：2026-11-10 16:15:51 浏览：9次
Online
报告开始：2026年11月21日 16:30（Asia/Kolkata）
报告时间：15min
所在会场：[S4] Artificial Intelligence and Intelligent Computing [S4-10] Artificial Intelligence and Intelligent Computing
暂无文件
摘要
Abstract—Composite risk scores that fuse supervised classifier
probabilities with unsupervised autoencoder reconstruction error
have recently been proposed to give fraud analysts a single
interpretable 0–100 triage signal. Existing composite-scoring
frameworks, however, combine these two signals through a single
fixed scalar weight tuned once by grid search, and explicitly leave
open whether a learned, context-aware fusion function would
perform better in practice. This paper implements and empir
ically evaluates that extension. We train a logistic-regression
meta-classifier over the outputs of a Random Forest classifier,
an XGBoost classifier, and an autoencoder anomaly detector
– augmented with transaction amount as a contextual feature
– and compare it against the fixed-weight composite baseline
on a held-out test partition preserving the extreme 0.172%
fraud rate of the ULB Credit Card Fraud benchmark. The
learned meta-fusion model strictly cost-dominates the standalone
Random Forest and XGBoost classifiers at every fraud-to-review
cost ratio, but does not uniformly dominate the fixed-weight
composite: a segment-level breakdown by transaction value shows
that meta-fusion wins decisively on low-value transactions while
the fixed-weight composite remains cost-optimal on mid- and
high-value transactions. Motivated by this finding, we propose a
segment-conditional deployment policy that applies meta-fusion
only to low-value transactions and the fixed-weight composite
elsewhere; this hybrid policy strictly dominates the fixed-weight
baseline and outperforms pure meta-fusion for essentially any
realistic cost ratio (r > 0.06). These results reframe fusion-layer
design not as a single global choice but as a segment-specific
operational decision, and provide financial institutions with a
concrete, reproducible protocol for deriving one.
Index Terms—Credit Card Fraud Detection, Score-Level Fu
sion, Meta-Classifier, Random Forest, XGBoo
关键词
Credit Card Fraud Detection, Machine Learning, Meta-Classifier, Score-Level Fusion, Random Forest, XGBoost, Autoencoder, Cost-Sensitive Learning, Risk Scoring, Anomaly Detection.
报告人
Aishwarya R
Student Anna University稿件作者
全部评论
重要日期
-
会议日期
11月20日
2026
至11月22日
2026
-
10月31日 2026
初稿截稿日期
-
11月07日 2026
提前注册日期
-
11月22日 2026
注册截止日期
主办单位
United Societies of Science
承办单位
Pondicherry University
协办单位
IEEE Section
IEEE Madras Section
IEEE Madras Section
联系方式
- ICon INDIA 2026
- ic******@usssociety.org
发表评论