DAO-CP

This project is a TensorLy implementation of DAO-CP: data-adaptive online CP decomposition (PLOS ONE).
This paper proposes a fast and accurate data-adaptive online CP factorization algorithm for high-order tensors.

Prerequisites

• Python 3.6+
• NumPy 1.19.1
• TensorLy 0.4.5
• Pandas 1.1.1
• SciPy 1.5.2
• wget 3.2

Usage

• Install all of the prerequisites with pip install -r requirements.txt.
• You can run the demo script by bash demo.sh <dataset_name>,
  which simply moves to src/ and runs python main.py <dataset_name>.
  • You can choose a dataset for demo (synthetic, video, stock, hall, korea).
• The script creates a tensor stream, performs online decomposition by various methods, and stores estimated tensors at out/.
  • Methods includes Full-CP, OnlineCP, DTD, and our proposed method DAO-CP.
  • Check decomposition results by loading .npy files.
• You can change the dataset and function parameters by modifying src/main.py.
• The core functions are implemented in src/dao-cp.

Datasets

Name	Order	Dimensions	Batch Sizes	Rank	L_s	L_r
synthetic	4	(1K, 10, 20, 30)	[10] * 100	30	1.2	1.1
video	4	(205, 240, 320, 3)	[5] * 41	30	6.0	2.0
stock	3	(3K, 140, 5)	[3] * 1K	20	6.0	5.0
hall	3	(200, 144, 176)	[10] * 20	20	0.5	0.1
korea	3	(10K, 323, 6)	[100] * 100	20	2.0	1.3

Reference

If you use this code, please cite the following paper.

@article{spc22dao,
	title={DAO-CP: Data-Adaptive Online CP decomposition for tensor stream},
	author={Son, Sangjun and Park, Yong-chan and Cho, Minyong and Kang, U},
	journal={PloS one},
	volume={17},
	number={4},
	pages={e0267091},
	year={2022},
	publisher={Public Library of Science San Francisco, CA USA}
}