An Adaptive Gradient Method for Online AUC Maximization

Authors: Yi Ding, Peilin Zhao, Steven Hoi, Yew-Soon Ong

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

Reproducibility Variable Result LLM Response
Research Type Experimental To demonstrate the effectiveness of the proposed algorithm, we analyze its theoretical bound, and further evaluate its empirical performance on both public benchmark datasets and anomaly detection datasets. We also empirically compared the proposed algorithm with several state-of-the-art online AUC optimization algorithms on both benchmark datasets and real-world online anomaly detection datasets.
Researcher Affiliation Academia 1School of Computer Engineering, Nanyang Technological University, 639798, Singapore 2Institute for Infocomm Research, 1 Fusionopolis Way, #21-01 Connexis, 138632, Singapore 3School of Information Systems, Singapore Management University, 178902, Singapore {ding0077,asysong}@ntu.edu.sg, zhaop@i2r.a-star.edu.sg, chhoi@smu.edu.sg
Pseudocode Yes Algorithm 1 The Ada OAM Algorithm and Algorithm 2 Adaptive Gradient Updating (AGU)
Open Source Code Yes The framework of experiments is based on the open-source library for large-scale online learning LIBOL 1 (Hoi, Wang, and Zhao 2014). 1http://libol.stevenhoi.org/
Open Datasets Yes All of these can be downloaded from LIBSVM 2 and UCI machine learning repository 3. 2http://www.csie.ntu.edu.tw/~cjlin/libsvmtools/ 3http://www.ics.uci.edu/~mlearn/MLRepository.html
Dataset Splits Yes Each dataset has been randomly divided into 5 folds, in which 4 folds are for training and the remaining fold is for testing. We also generate 4 independent 5-fold partitions per dataset to further reduce the variations. Therefore, the reported AUC value is an average of 20 runs for each dataset. 5-fold cross validation is conducted on the training sets to decide the learning rate η ∈ 2[−10:10] and the regularization parameter λ ∈ 2[−10:2].
Hardware Specification Yes All experiments were run with MATLAB on a computer workstation with 16GB memory and 3.20GHz CPU.
Software Dependencies No The paper mentions 'MATLAB' and the 'LIBOL' library but does not provide specific version numbers for these software components.
Experiment Setup Yes 5-fold cross validation is conducted on the training sets to decide the learning rate η ∈ 2[−10:10] and the regularization parameter λ ∈ 2[−10:2]. For OAMgra and OAMseq, the buffer size is fixed at 100 as suggested in (Zhao et al. 2011).