A Unified Framework for Rank-based Loss Minimization

Authors: Rufeng Xiao, Yuze Ge, Rujun Jiang, Yifan Yan

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments conducted on synthetic and real datasets illustrate the effectiveness and efficiency of the proposed algorithm.
Researcher Affiliation Academia Rufeng Xiao Yuze Ge Rujun Jiang Yifan Yan School of Data Science, Fudan University {rfxiao21,yzge23}@m.fudan.edu.cn rjjiang@fudan.edu.cn yanyf21@m.fudan.edu.cn
Pseudocode Yes Algorithm 1 ADMM framework. Algorithm 2 A refined pool-adjacent-violators algorithm for solving problem (10)
Open Source Code Yes The source code is available in the https://github.com/Rufeng Xiao/ADMM-for-rank-based-loss.
Open Datasets Yes We utilize the datasets.make_classification() function from the Python package scikit-learn [38] to generate two-class classification problem datasets of various dimensions. The SVMguide [26] dataset, frequently used in support vector machines, is included in our experiments. We also employ AD [29], which comprises potential advertisements for Internet pages, and Monks [50], the dataset based on the MONK s problem that served as the first international comparison of learning algorithms. The Splice dataset from the UCI [18] is used for the task of recognizing DNA sequences as exons or introns. We additionally include Australian , Phoneme , and Titanic dataset from [17]. Lastly, the UTKFace dataset [53] is used to predict gender based on facial images.
Dataset Splits Yes Consistent with their experimental setup, we randomly split each dataset into a 50% training set, a 25% validation set, and a 25% test set.
Hardware Specification Yes All algorithms are implemented in Python 3.8 and all the experiments are conducted on a Linux server with 256GB RAM and 96-core AMD EPYC 7402 2.8GHz CPU.
Software Dependencies No The paper states 'All algorithms are implemented in Python 3.8' but does not provide specific version numbers for other key software components or libraries used in the experiments.
Experiment Setup Yes In the experiments, we set the maximum iteration limit in our algorithm to 300. We set the γ in Section 4 to γk = max{10 5 0.9k, 10 9}. For simplicity, we set the regularization parameter µ as 0.01 and q as 0.8. Both SGD and LSVRG are run for 2000 epochs and the batch size of SGD is set to 64, and the epoch lengths are set to 100 as recommended in [36]. We set γ = 0.61 and δ = 0.69 in (5) based on the recommendations in [45]. We set B = log (1 + exp ( 5)) in (4).