Regret Bounds for Online Kernel Selection in Continuous Kernel Space

Authors: Xiao Zhang, Shizhong Liao, Jun Xu, Ji-Rong Wen10931-10938

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical studies verified the correctness of the theoretical regret analyses.
Researcher Affiliation Academia 1 Gaoling School of Articial Intelligence, Renmin University of China 2 Beijing Key Laboratory of Big Data Management and Analysis Methods 3 College of Intelligence and Computing, Tianjin University {zhangx89, junxu, jrwen}@ruc.edu.cn, szliao@tju.edu.cn
Pseudocode Yes Category 1: Online Kernel Selection by Selection-Post Training (OKS-SPT) and Category 2: Online Kernel Selection by Training-Post Selection (OKS-TPS) are presented as algorithm blocks with 'Require', 'Initialize', 'for', 'end for'.
Open Source Code No The paper does not provide any statement about releasing source code or links to a code repository for the described methodology.
Open Datasets Yes We merged the training set and testing set into one dataset for each benchmark dataset5. (Footnote 5: http://www.csie.ntu.edu.cn/~cjlin/libsvmtools/datasets/)
Dataset Splits No The paper states: 'We merged the training set and testing set into one dataset for each benchmark dataset' and 'there is no delineation among training, validation and testing phases in online learning', indicating that traditional validation splits are not provided or applicable in their online learning setting.
Hardware Specification Yes implemented in R 3.3.2 on a machine with 4-core Intel Core i7 3.60 GHz CPU and 16GB memory.
Software Dependencies Yes implemented in R 3.3.2
Experiment Setup Yes For all the algorithms, we used the hinge loss functions, tuned the stepsizes of online gradient descent in a range of 10[−5:+1:0], and selected the initial kernel σ1 in {2(i+1)/2, i = [−14 : +1 : 10]} uniform randomly, since small σ1 may lead to the vanishing of the gradients. In our categories, we set µ = 0.1 and used budgeted versions of the proposed categories that stop updating the buffer under a fixed budget B = 200. Besides, we set the dimension of random features D = 400 in RRF, and set the smoothing parameter and stepsize of OKS as in (Yang et al. 2012).