• Login
    View Item 
    •   DSpace@RPI Home
    • Rensselaer Libraries
    • RPI Theses Online (Complete)
    • View Item
    •   DSpace@RPI Home
    • Rensselaer Libraries
    • RPI Theses Online (Complete)
    • View Item
    JavaScript is disabled for your browser. Some features of this site may not work without it.

    Matrix sampling algorithms for topics in machine learning

    Author
    Paul, Saurabh
    View/Open
    175999_Paul_rpi_0185E_10586.pdf (1.167Mb)
    Other Contributors
    Drineas, Petros; Magdon-Ismail, Malik; Zaki, Mohammed J., 1971-; Bennett, Kristin P.;
    Date Issued
    2015-05
    Subject
    Computer science
    Degree
    PhD;
    Terms of Use
    This electronic version is a licensed copy owned by Rensselaer Polytechnic Institute, Troy, NY. Copyright of original work retained by author.;
    Metadata
    Show full item record
    URI
    https://hdl.handle.net/20.500.13015/1476
    Abstract
    We study the application of matrix sampling algorithms on four problems in machine learning, namely: (i) Feature Selection for Linear Support Vector Machines, (ii) Feature Selection for Ridge Regression, (iii) Core-set Construction for Canonical Correlation Analysis and (iv) Adaptive Sampling algorithm for matrix reconstruction. We provide both theoretical performance guarantees and empirical evidence to indicate the effectiveness of our methods. A more detailed description is given below.; 4. Adaptive Sampling algorithm for matrix reconstruction: We introduce a new adaptive sampling algorithm for computing low-rank matrix approximation. We are given a matrix A and a target rank k. The algorithm runs in t iterations and selects a subset of columns of the matrix. It computes a rank-k approximation to the matrix that is as good as the best rank-k approximation that would have been obtained by using all the columns.; 3. Core-set Construction for Canonical Correlation Analysis (CCA): We introduce two algorithms for core-set construction from CCA. The algorithms select a subset of data-points from the pair of matrices and compute approximations to all canonical correlations with provable guarantees. We show that any set of canonical weights of sampled pair of matrices can be used to obtain a set of approximately orthogonal canonical vectors of the original pair of matrices.; 2. Feature Selection for Ridge Regression: We introduce an unsupervised feature selection technique for regularized least squares classification (RLSC), (which is the classification analogue to ridge regression) and provide worst-case guarantees of the generalization power of the classification function after feature selection with respect to the classification function obtained using all features. We provide risk bounds for feature selection methods for ridge regression in the fixed design setting.; 1. Feature Selection for Linear Support vector Machines: We present feature selection algorithms for linear Support Vector Machines (SVM), which can be used in an unsupervised or supervised setting. We prove that the margin in the feature space is preserved to within $\epsilon$-relative error of the margin in the full feature space in the supervised setting. In the unsupervised setting, we also provide worst-case guarantees of the radius of the minimum enclosing ball, thereby ensuring comparable generalization as in the full feature space. We also present feature extraction techniques for linear SVM, which preserve both margin and data-radius, upto $\epsilon$-relative error.;
    Description
    May 2015; School of Science
    Department
    Dept. of Computer Science;
    Publisher
    Rensselaer Polytechnic Institute, Troy, NY
    Relationships
    Rensselaer Theses and Dissertations Online Collection;
    Access
    Restricted to current Rensselaer faculty, staff and students. Access inquiries may be directed to the Rensselaer Libraries.;
    Collections
    • RPI Theses Online (Complete)

    Browse

    All of DSpace@RPICommunities & CollectionsBy Issue DateAuthorsTitlesSubjectsThis CollectionBy Issue DateAuthorsTitlesSubjects

    My Account

    Login

    DSpace software copyright © 2002-2022  DuraSpace
    Contact Us | Send Feedback
    DSpace Express is a service operated by 
    Atmire NV