notesum.ai

Published at November 5

DiffLM: Controllable Synthetic Data Generation via Diffusion Language Models

cs.LG

cs.AI

cs.CL

Released Date: November 5, 2024

Authors: Ying Zhou, Xinyao Wang, Yulei Niu, Yaojie Shen, Lexin Tang, Fan Chen, Ben He, Le Sun, Longyin Wen

Arxiv: http://arxiv.org/abs/2411.03250v1

Refer to caption

Method	Adult		Default		Magic		Shoppers		Beijing
Method	MLE $\uparrow$	$\rho$ $\downarrow$	MLE $\uparrow$	$\rho$ $\downarrow$	MLE $\uparrow$	$\rho$ $\downarrow$	MLE $\uparrow$	$\rho$ $\downarrow$	MLE $\downarrow$	$\rho$ $\downarrow$
Real	0.927	-	0.770	-	0.946	-	0.926	-	0.423	-
SMOTE	0.899	1.60	0.741	1.48	0.934	0.91	0.911	2.68	0.593	1.85
CTGAN	0.886	16.84	0.696	16.83	0.855	9.810	0.875	21.15	0.902	21.39
TVAE	0.878	14.22	0.724	10.17	0.887	8.250	0.871	24.51	0.770	19.16
GOGGLE	0.778	16.97	0.584	17.02	0.654	1.900	0.658	22.33	1.090	16.93
CoDi	0.871	21.38	0.525	15.77	0.932	11.56	0.865	31.84	0.818	16.94
TabSyn	0.915	0.58	0.764	0.85	0.938	0.88	0.920	1.43	0.582	1.12
GReaT	0.913	12.12	0.755	19.94	0.888	16.16	0.902	14.51	0.653	8.25
DiffLM	0.894	9.16	0.793	9.33	0.910	7.04	0.912	14.43	0.717	6.05