Diverse Power Iteration Embeddings: Theory and Practice
Abstract
Manifold learning, especially spectral embedding, is known as one of the most effective learning approaches on high dimensional data, but for real-world applications it raises a serious computational burden in constructing spectral embeddings for large datasets. To overcome this computational complexity, we propose a novel efficient embedding construction, Diverse Power Iteration Embedding (DPIE). DPIE shows almost the same effectiveness of spectral embeddings and yet is three order of magnitude faster than spectral embeddings computed from eigen-decomposition. Our DPIE is unique in that (1) it finds linearly independent embeddings and thus shows diverse aspects of dataset; (2) the proposed regularized DPIE is effective if we need many embeddings; (3) we show how to efficiently orthogonalize DPIE if one needs; and (4) Diverse Power Iteration Value (DPIV) provides the importance of each DPIE like an eigen value. As a result, such various aspects of DPIE and DPIV ensure that our algorithm is easy to apply to various applications, and we also show the effectiveness and efficiency of DPIE on clustering, anomaly detection, and feature selection as our case studies.
- Authors:
- Publication Date:
- Research Org.:
- Brookhaven National Lab. (BNL), Upton, NY (United States)
- Sponsoring Org.:
- USDOE; USDOE Office of Science (SC), Advanced Scientific Computing Research (SC-21)
- OSTI Identifier:
- 1327442
- Alternate Identifier(s):
- OSTI ID: 1327443; OSTI ID: 1345738
- Report Number(s):
- BNL-113578-2017-JA
Journal ID: ISSN 1041-4347; 7322265
- Grant/Contract Number:
- PD 15-025; SC0003361; SC00112704
- Resource Type:
- Published Article
- Journal Name:
- IEEE Transactions on Knowledge and Data Engineering
- Additional Journal Information:
- Journal Name: IEEE Transactions on Knowledge and Data Engineering Journal Volume: 28 Journal Issue: 10; Journal ID: ISSN 1041-4347
- Publisher:
- Institute of Electrical and Electronics Engineers
- Country of Publication:
- United States
- Language:
- English
- Subject:
- 97 MATHEMATICS AND COMPUTING; power iteration; approximated spectral analysis
Citation Formats
Huang, Hao, Yoo, Shinjae, Yu, Dantong, and Qin, Hong. Diverse Power Iteration Embeddings: Theory and Practice. United States: N. p., 2016.
Web. doi:10.1109/TKDE.2015.2499184.
Huang, Hao, Yoo, Shinjae, Yu, Dantong, & Qin, Hong. Diverse Power Iteration Embeddings: Theory and Practice. United States. https://doi.org/10.1109/TKDE.2015.2499184
Huang, Hao, Yoo, Shinjae, Yu, Dantong, and Qin, Hong. Sat .
"Diverse Power Iteration Embeddings: Theory and Practice". United States. https://doi.org/10.1109/TKDE.2015.2499184.
@article{osti_1327442,
title = {Diverse Power Iteration Embeddings: Theory and Practice},
author = {Huang, Hao and Yoo, Shinjae and Yu, Dantong and Qin, Hong},
abstractNote = {Manifold learning, especially spectral embedding, is known as one of the most effective learning approaches on high dimensional data, but for real-world applications it raises a serious computational burden in constructing spectral embeddings for large datasets. To overcome this computational complexity, we propose a novel efficient embedding construction, Diverse Power Iteration Embedding (DPIE). DPIE shows almost the same effectiveness of spectral embeddings and yet is three order of magnitude faster than spectral embeddings computed from eigen-decomposition. Our DPIE is unique in that (1) it finds linearly independent embeddings and thus shows diverse aspects of dataset; (2) the proposed regularized DPIE is effective if we need many embeddings; (3) we show how to efficiently orthogonalize DPIE if one needs; and (4) Diverse Power Iteration Value (DPIV) provides the importance of each DPIE like an eigen value. As a result, such various aspects of DPIE and DPIV ensure that our algorithm is easy to apply to various applications, and we also show the effectiveness and efficiency of DPIE on clustering, anomaly detection, and feature selection as our case studies.},
doi = {10.1109/TKDE.2015.2499184},
journal = {IEEE Transactions on Knowledge and Data Engineering},
number = 10,
volume = 28,
place = {United States},
year = {Sat Oct 01 00:00:00 EDT 2016},
month = {Sat Oct 01 00:00:00 EDT 2016}
}
https://doi.org/10.1109/TKDE.2015.2499184
Web of Science