High-dimensional Similarity Learning via Dual-sparse Random Projection

Authors: Dezhong Yao, Peilin Zhao, Tuan-Anh Nguyen Pham, Gao Cong

IJCAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we first present our study on Du RPRSL for ranking and classification tasks. Then we give a case study on the support of dual-sparse regularized Du RPRSL.
Researcher Affiliation Collaboration Dezhong Yao1, Peilin Zhao2, Tuan-Anh Nguyen Pham1 and Gao Cong3 1 Rolls-Royce@NTU Corporate Lab, Nanyang Technological University, Singapore 2 South China University of Technology; Tencent AI Lab, China 3 School of Computer Science and Engineering, Nanyang Technological University, Singapore
Pseudocode Yes Algorithm 1 Dual Random Projection Method for Relative Similarity Learning (Du RPRSL) and Algorithm 2 SDCA: Stochastic Dual Coordinate Ascent for Relative Similarity Learning
Open Source Code No The paper does not provide any concrete access information (e.g., specific repository link, explicit code release statement, or code in supplementary materials) for the methodology described in this paper.
Open Datasets Yes To examine the effectiveness of the proposed method, we tested on six public datasets: Protein, Gisette, RCV1, URL, Caltech256, and BBC from LIBSVM, Caltech, and UCD, as shown in Table 1.
Dataset Splits Yes For evaluation, we used the standard training and testing split given by the providers, except for Caltch30 and BBC. For these two datasets, we randomly split them into a training set (70%) and a test set (30%). To generate a triplet (xt, x+ t , x t ), xt is firstly randomly selected from the whole training set, then x+ t is randomly selected from the subset of training set, which consists of the examples with the same class of xt, at last, x t is randomly selected from the rest of training set, which consists of the examples with different classes of xt. To make a fair comparison, all methods adopted the same experimental setup. We randomly selected T=50, 000 triplets as training instances and set the number of epochs to be 5 for all stochastic methods. The average results over five trials were reported finally. Cross-validation was used to select the values of hyperparameters for all algorithms.
Hardware Specification No The paper does not provide specific hardware details (e.g., exact GPU/CPU models, processor types with speeds, memory amounts, or detailed computer specifications) used for running its experiments.
Software Dependencies No The paper does not provide specific ancillary software details (e.g., library or solver names with version numbers) needed to replicate the experiment.
Experiment Setup Yes To make a fair comparison, all methods adopted the same experimental setup. We randomly selected T=50, 000 triplets as training instances and set the number of epochs to be 5 for all stochastic methods. The average results over five trials were reported finally. Cross-validation was used to select the values of hyperparameters for all algorithms. Specifically, the parameters set by cross-validation included: the aggressiveness parameter C for OASIS (C {1, 0.1, 0.08, ..., 0.01}) and λ {5e-2, 5e-1, ..., 5e+6}. Moreover, the hinge loss was used in the implementation of the proposed algorithms.