notesum.ai

Published at December 6

cs.LG

I.2

Released Date: December 6, 2024

Authors: Margaux Tornqvist¹, Jean-Daniel Zucker², Tristan Fauvel¹, Nicolas Lambert¹, Mathilde Berthelot¹, Antoine Movschin¹

Aff.: ¹Quinten Health; ²IRD, Sorbonne University

SDG Model	$D_{train}$ (Ref.)	GPT-4 (Ours)	GC	CTGAN	TVAE
Fidelity	$D_{Test}$	$D_{Test}$	$D_{Test}$	$D_{Test}$	$D_{Test}$
Column Shapes ( $\uparrow$ )	0.956±0.017	0.778±0.008	0.913±0.004	0.839±0.021	0.885±0.011
Column Pair Trends ( $\uparrow$ )	0.914±0.014	0.807±0.010	0.886±0.007	0.851±0.009	0.879±0.012
KSComplement ( $\uparrow$ )	0.950±0.018	0.767±0.011	0.902±0.002	0.814±0.024	0.887±0.011
TVComplement ( $\uparrow$ )	0.978±0.014	0.820±0.012	0.951±0.017	0.930±0.031	0.877±0.031
CorrelationSimilarity ( $\uparrow$ )	0.971±0.010	0.947±0.005	0.969±0.004	0.917±0.007	0.946±0.009
ContingencySimilarity ( $\uparrow$ )	0.961±0.018	0.739±0.017	0.916±0.019	0.885±0.043	0.801±0.044
WD ( $\downarrow$ )	0.476±0.182	2.274 ±0.126	1.023±0.077	2.14±0.340	1.009±0.144
JSD ( $\uparrow$ )	0.183±0.007	0.154±0.003	0.195±0.004	0.219±0.007	0.173±0.004
LogisticDetection ( $\uparrow$ )	0.968±0.039	0.398±0.022	0.866±0.040	0.405±0.118	0.553±0.089
Privacy	$D_{Train}$	$D_{Train}$	$D_{Train}$	$D_{Train}$
NewRowSynthesis	-	1±0	1±0	1±0	1±0
DCR (5th p.) ( $\uparrow$ )	-	1.556±0.038	1.347±0.030	1.391±0.042	1.192±0.048
NNDR (5th p.) ( $\uparrow$ )	-	0.765±0.012	0.727±0.016	0.737±0.015	0.768±0.016
CategoricalCAP ( $\uparrow$ )	-	0.882±0.013	0.911±0.010	0.897±0.015	0.857±0.017
Utility	$D_{Test}$	$D_{Test}$	$D_{Test}$	$D_{Test}$	$D_{Test}$
TSTR (F1 score) ( $\uparrow$ )	0.978±0.112	0.979±0.023	0.960±0.014	0.838±0.117	0.953±0.044
TATR (F1 score) ( $\uparrow$ )	-	0.983±0.018	0.969±0.008	0.883±0.093	0.960±0.036
SDG time	-	$\sim 1:30hrs$	$\sim 1s$	$\sim 3min$	$\sim 1min$