AUC Maximization under Positive Distribution Shift

Authors: Atsutoshi Kumagai, Tomoharu Iwata, Hiroshi Takahashi, Taishi Nishiyama, Yasuhiro Fujiwara

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The effectiveness of the proposed method is experimentally shown with six real-world datasets.
Researcher Affiliation Industry Atsutoshi Kumagai NTT atsutoshi.kumagai@ntt.comTomoharu Iwata NTT tomoharu.iwata@ntt.comHiroshi Takahashi NTT hiroshibm.takahashi@ntt.comTaishi Nishiyama NTT Security Holdings, NTT taishi.nishiyama@security.nttYasuhiro Fujiwara NTT yasuhiro.fujiwara@ntt.com
Pseudocode Yes Algorithm 1 Training procedure of the proposed method
Open Source Code No The code is proprietary.
Open Datasets Yes We utilized four widely used real-world datasets in the main paper: MNIST [30], Fashion MNIST [57], SVHN [40], and CIFAR10 [26]. ...we also evaluated the proposed method with two tabular datasets with distribution shifts (HReadmission and Hypertension) [14].
Dataset Splits Yes For each dataset, we used 10 positive and 5, 000 unlabeled data in the training distribution and 5, 000 unlabeled data in the test distribution for training. In addition, we used 5 positive and 500 unlabeled data in the training distribution and 500 unlabeled data in the test distribution for validation. We used 1, 500 positive and 1, 500 negative data in the test distribution as test data for evaluation.
Hardware Specification Yes All methods were implemented using Pytorch [43] and all experiments were conducted on a Linux server with an Intel Xeon CPU and A100 GPU.
Software Dependencies No All methods were implemented using Pytorch [43]... The paper mentions a software library (Pytorch) but does not specify a version number or other software dependencies with specific version numbers.
Experiment Setup Yes For all methods, we used the Adam optimizer [24] with a learning rate of 10 4. We set a mini-batch size M to 512, a positive mini-batch size P to 10, and the maximum number of epochs to 200. The loss on validation data was used for early stopping to avoid overfitting.