On the consistent estimation of optimal Receiver Operating Characteristic (ROC) curve

Authors: Renxiong Liu, Yunzhang Zhu

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we examine the operating characteristics of the three methods analyzed in this article with both simulation studies and a real data example. In our simulation, we investigate the effect of model mis-specification by considering linear classifiers over two simulated data sets. One dataset is generated under a linear discriminant analysis (LDA) setting, which imitates the scenario where model is correctly specified. The other is based on a quadratic discriminant analysis (QDA) setting, which is used to show the differences across three methods under model mis-specification. We also compare the three methods by using a bank marketing data set [Moro et al., 2014], which allows us to investigate the effect model mis-specification on the performance of ROC curve estimation methods in real problem setting. Throughout all the experiments, we consider linear classification methods including constrained ψ-learning, weighted SVM and its cutoff version, and also typical kernel methods that include kernel weighted SVM and its cutoff version
Researcher Affiliation Academia Renxiong Liu Department of Statistics Ohio State University Columbus, OH 43210 liu.6732@buckeyemail.osu.edu Yunzhang Zhu Department of Statistics Ohio State University Columbus, OH 43210 zhu.219@osu.edu
Pseudocode No No pseudocode or algorithm blocks found in the paper.
Open Source Code No [No] We will release our codes later.
Open Datasets Yes Example 3 (Real data example). This example considers a bank marketing dataset [Moro et al., 2014], which records the direct marketing campaigns of a Portuguese banking institution and are available at https://archive.ics.uci.edu/ml/datasets/Bank+Marketing.
Dataset Splits No No explicit train/validation/test dataset splits by percentage or count are provided. The paper mentions
Hardware Specification No No specific hardware details (GPU/CPU models, memory, etc.) for running experiments are provided.
Software Dependencies No No specific ancillary software details with version numbers are provided.
Experiment Setup Yes For all three examples, to generate the estimated ROC curve we vary weight w and the constraint upper bound α from {i/500 | i = 0, 1, . . . , 500} for the weighted method and the constrained method, respectively. For the two cutoff methods, we choose w = 1/2.