notesum.ai

Published at November 17

TabDeco: A Comprehensive Contrastive Framework for Decoupled Representations in Tabular Data

cs.AI

Released Date: November 17, 2024

Authors: Suiyao Chen¹, Jing Wu², Yunxiao Wang¹, Cheng Ji¹, Tianpei Xie¹, Daniel Cociorva¹, Michael Sharps², Cecile Levasseur¹, Hakan Brunzell¹

Aff.: ¹Buyer Abuse Prevention; ²WWPS Solution Architecture

Arxiv: http://arxiv.org/abs/2411.11148v1

Dataset size	45,211	7,043	12,330	32,561	17,898	10,000	4,601	1,055	495,141	58,310	60,000
Feature size	16	27	17	14	8	11	57	41	49	147	784
Method/Dataset	Bank	Blastchar	Shoppers	Income	HTRU2	Shrutime	Spambase	QSARBio	Forest	Volkert $\dagger$	MNIST $\dagger$
Logistic Reg.	90.73	82.34	87.03	92.12	98.23	83.37	92.77	84.06	84.79	53.87	89.89
Random Forest	89.12	80.63	89.87	88.04	96.41	80.87	98.02	91.49	98.80	66.25	93.75
XGBoost	92.96	81.78	92.51	92.31	97.81	83.59	98.91	92.70	95.53	68.95	94.13
LightGBM	93.39	83.17	93.20	92.57	98.10	85.36	99.01	92.97	93.29	67.91	95.20
CatBoost	90.47	84.77	93.12	90.80	97.85	85.44	98.47	93.05	85.36	66.37	96.60
MLP	91.47	59.63	84.71	92.08	98.35	73.70	66.74	79.66	96.81	63.02	93.87
VIME	76.64	50.08	74.37	88.98	97.02	70.24	69.24	81.04	75.06	64.28	95.77
TabNet	91.76	79.61	91.38	90.72	97.58	75.24	97.93	67.55	96.37	56.83	96.79
TabTransormer	91.34	81.67	92.70	90.60	96.56	85.60	98.50	91.80	84.96	57.98	88.74
SwitchTab	91.80	83.56	91.20	89.79	97.15	84.89	96.06	93.85	97.66	68.90	96.80
SAINT	93.12	84.58	93.22	90.99	98.28	85.79	97.66	93.34	99.06	68.87	97.48
SAINT-s	92.84	83.57	92.31	90.59	98.58	85.43	97.83	92.48	98.85	69.61	97.45
SAINT-f	92.68	85.18	93.10	91.23	98.54	85.09	97.40	91.91	98.96	58.76	91.25
TabDeco	93.34	84.96	93.53	91.17	98.57	86.14	97.79	93.80	99.15	69.46	97.85
TabDeco-s	93.00	84.16	92.66	90.95	98.60	85.93	97.97	94.87	99.06	69.91	97.71
TabDeco-f	92.48	85.00	93.05	91.26	98.60	85.00	97.36	92.08	99.00	58.52	95.56