notesum.ai

Published at November 9

A Picture is Worth A Thousand Numbers: Enabling LLMs Reason about Time Series via Visualization

cs.LG

cs.AI

Released Date: November 9, 2024

Authors: Haoxin Liu¹, Chenghao Liu², B. Aditya Prakash¹

Aff.: ¹Georgia Institute of Technology; ²Salesforce Research Asia

Arxiv: http://arxiv.org/abs/2411.06018v1

Refer to caption

Reasoning Pattern		Simple Deterministic		Complex Deterministic		Probabilistic
Reasoning Task		RCW	TEE	ECG	EMG	CTU	HAR
Metric		ACC(%)	ACC(%)	ACC(%)	ACC(%)	ACC(%)	ACC(%)
Random Guessing		50.00	14.29	25.00	33.33	50.00	16.67
Supervised Time-series Models (8)	Transformer	64.12	59.52	25.00	86.67	59.20	87.26
	Autoformer	62.59	26.19	23.95	46.67	67.20	75.04
	Informer	75.51	59.52	22.39	66.66	67.20	85.83
	FEDformer	76.59	42.86	26.40	73.33	51.60	89.88
	PatchTST	82.11	57.14	24.82	60.00	64.00	79.60
	iTransformer	76.92	21.43	24.48	46.67	46.40	89.49
	TimesNet	80.23	61.90	26.20	73.33	64.00	88.65
	DLinear	56.96	47.63	23.61	46.67	52.40	48.97
Zero-shot LLMs	GPT-4o (numeric)	50.00	21.43	25.00	33.33	45.45	29.17
	GPT-4o (VL-Time)	70.02	24.88	26.33	33.33	50.71	37.50
	Improvement	+40.04%	+16.10%	+5.32 %	+0.00%	+11.57%	+28.56%
	Win Supervised	3/8	1/8	7/8	0/8	1/8	0/8
Few-shot ICL LLMs	GPT-4o (numeric)	50.00	35.71	31.25	33.33	50.00	12.50
	GPT-4o (VL-Time)	91.03	64.29	43.75	91.67	63.64	66.67
	Improvement	+82.06%	+80.03%	+40.00%	+175.04%	+27.28%	+433.36%
	Win Supervised	8/8	8/8	8/8	8/8	4/8	1/8