Classification with Low Rank and Missing Data

Authors: Elad Hazan, Roi Livni, Yishay Mansour

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

Reproducibility Variable Result LLM Response
Research Type Experimental We complement our theoretical contributions with experimental findings that show superior classification performance both on synthetic data and on publicly-available recommendation data.
Researcher Affiliation Collaboration Elad Hazan EHAZAN@CS.PRINCETON.EDU Princeton University and Microsoft Research, Herzliya Roi Livni ROI.LIVNI@MAIL.HUJI.AC.IL The Hebrew University of Jerusalem and Microsoft Research, Herzliya Yishay Mansour MANSOUR.YISHAY@GMAIL.COM Microsoft Research, Hertzelia and Tel Aviv University
Pseudocode Yes Algorithm 1 KARMA: Kernelized Algorithm for Riskminimization with Missing Attributes
Open Source Code No No explicit statement regarding the release of source code or a link to a code repository was found.
Open Datasets Yes The Jester data set was collected from (Goldberg et al., 2001), The books dataset was collected from (Ziegler et al., May 1014, 2005) and we ve also used the Movie Lens data set2. 2available here: http://grouplens.org/datasets/ movielens/
Dataset Splits Yes We chose γ and λ using a holdout set.
Hardware Specification No No specific hardware details (e.g., GPU models, CPU types, memory amounts) used for running experiments were provided.
Software Dependencies No No specific software dependencies with version numbers (e.g., 'Python 3.8', 'PyTorch 1.9') were provided for the implementation or experiments.
Experiment Setup Yes We evaluated the loss with γ = {1, 2, 3, 4} and C = {10 5, 10 4, . . . , 104, 105}. We chose γ and λ using a holdout set. A constant feature was added to allow bias and the data was normalized. For binary classification we let ℓbe the Hinge loss, for multiclass we used the multiclass Hinge loss as described in (Crammer & Singer, 2002) and finally for regression tasks we used squared loss (which was also used at test time).