notesum.ai

Published at November 6

UniTraj: Universal Human Trajectory Modeling from Billion-Scale Worldwide Traces

cs.ET

cs.AI

cs.LG

cs.SI

physics.soc-ph

Released Date: November 6, 2024

Authors: Yuanshao Zhu¹, James Jianqiao Yu², Xiangyu Zhao³, Xuetao Wei¹, Yuxuan Liang⁴

Aff.: ¹Southern University of Science and Technology; ²University of York; ³City University of Hong Kong; ⁴The Hong Kong University of Science and Technology (Guangzhou)

Arxiv: http://arxiv.org/abs/2411.03859v1

Methods	WorldTrace		Chengdu		Xi’an		GeoLife		Grab-Posisi		Porto
Methods	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE
Linear	427.68	516.15	205.74	258.52	176.49	220.87	196.85	249.76	507.41	617.28	396.61	482.39
DHTR	220.35	302.47	75.19	98.68	62.85	83.43	80.04	168.25	351.20	415.16	194.37	232.59
Transformer	130.82	147.62	55.23	62.85	45.85	51.96	94.68	113.77	136.58	163.29	104.36	126.96
DeepMove	51.16	62.29	29.32	39.02	27.31	35.67	86.38	107.78	126.93	168.07	136.66	174.96
TrajBERT	58.13	70.14	26.48	33.83	19.45	25.13	34.53	43.24	112.68	136.24	78.77	99.23
TrajFM	47.64	58.92	19.10	25.09	18.86	24.13	59.34	64.24	107.64	130.69	71.15	92.96
UniTraj (zero-shot)	10.22	13.56	11.98	20.94	8.93	13.83	37.21	63.89	114.07	167.01	78.28	100.14
Improvement(%)	^↑78.55	^↑76.99	^↑37.28	^↑16.54	^↑52.65	^↑42.69	_↓7.76	_↓47.46	_↓5.97	_↓27.79	_↓10.02	_↓7.72
UniTraj (fine-tune)	6.94	9.67	6.92	10.41	6.50	9.93	23.23	34.70	48.95	69.23	60.18	79.76
Improvement(%)	^↑85.43	^↑83.59	^↑63.77	^↑58.51	^↑65.54	^↑58.85	^↑32.73	^↑19.75	^↑54.52	^↑47.03	^↑15.42	^↑14.20