Dissertation > Excellent graduate degree dissertation topics show

Lanczos Bidiagonalization Based Initialization: A Fast Head Start for Nonnegative Matrix Factorization

Author: WangZuoSheng
Tutor: LuLinZuo
School: Xiamen University
Course: Computational Mathematics
Keywords: Lanczos bidiagonalization Nonnegative matrix factorization Singular value decomposition Low-rank approxiamtion
CLC: TP391.41
Type: Master's thesis
Year: 2009
Downloads: 120
Quote: 0
Read: Download Dissertation

Abstract


In 1999,"Nature" published a paper showing an outstanding result in the research of non-negative matrix of mathematics,written by Professor Lee and Professor Seung.In this paper,the authors present a new method of matrix decomposition,named non-negative matrix factorization(NMF) algorithm, a method applied on the condition that all the elements in the matrix are non-negative.Because of the broad application of non-negative matrix factorization in real life,the publication of this paper caught the eyes of a number of researchers studying in some specific field for the following reasons:On the one hand,many large-scale data analysis in scientific research needs to be processed effectively in the form of matrixes,and NMF can provide a new way for human in processing the large-scale data;On the other hand,compared with some of other traditional algorithms,NMF decomposition algorithm shows a number of advantages,such as the simplicity in realization,easily explainable toward the form of and the result of decomposition,as well as less storage space.In[2],C.Boutsidisa and E.Gallopoulos proposed a SVD based initialization algorithm designed to enhance the initialization stage of nonnegativc matrix factorization.Combined with other NMF algorithms,it can accelerate convergenece rate rapidlly.For a large nonnegativc matrix,a Lanczos bidiago nalization process is utilized to obtain a nonnegativc bidiagonal matrix of low rank,and then every unit rank matrix produced from the Lanczos process is approximated by its nonncgative section in the same way as developed in[2]. This results in a noval initialization algorithm for vomiegative matrix factorization (NMF).It can readily be combined with existing NMF algorithms and may contain a little randomization.Some numerical experiments demonstrate that the new initialization algorithm is more efficient than the SVD based initialization algorithm presented in[2].Five sections are folded in this thesis.Section 1 give us the introduction of NMF and the initializaiton of NMF and looks back the former works of this area.In section2,we review the Lanczos bidiagonalization process to get a low-rank approximation.In sectin3,we present a strategy to get nonnegative approximation matrices W and H from the Lanczos bidiagonalization process. In section4,we give some experiments to compare our method with C.Boutsidis and E.Gallopoulous’ method and demonstrate the advantages of our methods. In section 5,we give a conclusion of our thesis.

Related Dissertations

  1. Research on Approaches of the Subjective Automated Assessment,TP391.1
  2. Collaborative Filtering Algorithm Research Based on Page Interest Degree,TP393.092
  3. A Study on the Theorem of Multi-wavelets and Its Construction Algorithm,O174.2
  4. Based on singular value decomposition Radon domain multiple suppression Method,P631.44
  5. Based on the empty block frequency domain combined dual watermark,TP309.7
  6. Based on the discrete cosine transform and wavelet domain image watermarking algorithm,TP309.7
  7. Several Classes of Constrained Matrix Equations and Their Minimum Rank Solutions,O241.6
  8. Digital watermarking key technology research,TP309.7
  9. Parameters Estimate for a Class of Linear Mixed Models,O212.1
  10. Research of Digital Image Watermarking Techniques,TP309.7
  11. Research on Subspace Analysis Based Face Recognition Algorithm,TP391.41
  12. The literature of resources for research and application clustering system,TP391.1
  13. Research of Encryption Image Digital Watermark Algorithm Based on Transform Domain,TP309.7
  14. Quaternion Matrix Singular Value Decomposition and Its Application,O151.23
  15. Acoustic Temperature Measurement Technique and Temperature Field Reconstruction Algorithm,TH811
  16. An Investigation of Theory and Application of New Methods on Depth Profile Recognition Based on Photoacoustic and Photothermal Detect Technology,TB303
  17. Algorithm for Weak and Small Target Detection under Complicated Background,TP391.41
  18. Research of Beamforming Technology Based on LTE Downlink,TN929.5
  19. Image watermarking algorithm based on wavelet transform and singular value decomposition,TP309.7
  20. Personalized Recommender System Based on Latent Semantic Analysis,TP391.1

CLC: > Industrial Technology > Automation technology,computer technology > Computing technology,computer technology > Computer applications > Information processing (information processing) > Pattern Recognition and devices > Image recognition device
© 2012 www.DissertationTopic.Net  Mobile