Attribute Efficient Linear Regression with Distribution-Dependent Sampling

Authors: Doron Kukliansky, Ohad Shamir

ICML 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We complement our theoretical analysis with experiments which support our claims. We conducted two sets of experiments: One on artificial data, which allows us to easily control data properties such as ED x2 1 2 and ED x2 1; And the other on a subset of the popular MNIST (Le Cun et al., 1998) data set, similar to (Cesa Bianchi et al., 2011; Hazan & Koren, 2012).
Researcher Affiliation Academia Doron Kukliansky DORON.KUKLIANSKY@WEIZMANN.AC.IL Ohad Shamir OHAD.SHAMIR@WEIZMANN.AC.IL Department of Computer Science and Applied Mathematics, Weizmann Institute of Science, Rehovot, Israel
Pseudocode Yes Algorithm 1 GAERR Parameters: B, η > 0 and qi for i [d] Input: training set S = {(xt, yt)}t [m] and k > 0 Output: regressor ˉw with ˉw 2 B
Open Source Code No The paper does not provide concrete access to source code for the methodology described.
Open Datasets Yes We conducted two sets of experiments: One on artificial data... And the other on a subset of the popular MNIST (Le Cun et al., 1998) data set, similar to (Cesa Bianchi et al., 2011; Hazan & Koren, 2012).
Dataset Splits Yes Similar to (Cesa-Bianchi et al., 2011; Hazan & Koren, 2012), we used 10-fold cross validation to optimize the parameters for each phase. For the two-phased algorithms, we set m1 = m 10, m2 = 9m 10
Hardware Specification No The paper does not provide specific details about the hardware used to run the experiments.
Software Dependencies No The paper does not provide specific software dependencies with version numbers.
Experiment Setup Yes Similar to (Cesa-Bianchi et al., 2011; Hazan & Koren, 2012), we used 10-fold cross validation to optimize the parameters for each phase. For the two-phased algorithms, we set m1 = m 10, m2 = 9m 10 , and run the AERR/AELR algorithm during the first phase, using its result as a starting point for the second phase. Unlike the theoretical analysis, we set ϵ to 0, since the theoretical upper confidence bound is conservative, and split the attribute budget evenly between the data point estimation and the inner product estimation as we found that these improve the empirical results.