TabLog: Test-Time Adaptation for Tabular Data Using Logic Rules

Authors: Weijieying Ren, Xiaoting Li, Huiyuan Chen, Vineeth Rakesh, Zhuoyi Wang, Mahashweta Das, Vasant G Honavar

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

Reproducibility Variable Result LLM Response
Research Type Experimental We present results of experiments using several benchmark data sets that demonstrate Tab Log is competitive with or improves upon the state-of-the-art methods for testtime adaptation of predictive models trained on tabular data.
Researcher Affiliation Collaboration 1Artificial Intelligence Research Laboratory, Center for Artificial Intelligence Foundations and Scientific Applications, Institute for Computational and Data Sciences, College of Information Sciences and Technology, The Pennsylvania State University, PA, United States. 2Visa Research, CA, United States 3College of Information Sciences and Technology, The Pennsylvania State University, PA, United States.
Pseudocode No The paper describes the Tab Log algorithm narratively and with a block diagram (Fig. 1), but does not include structured pseudocode or an algorithm block.
Open Source Code Yes Our code is available at https:// github.com/Weijieying Ren/Tab Log.
Open Datasets Yes Our experiments used (i) four publicly available benchmark data sets that exhibit natural distribution shifts: ASSISTments, Sepsis, Hospital Readmission, and Physio Net, available as part of the Table Shift benchmark (Gardner et al., 2023); and (ii) four tabular data sets Airbnb, Channel, Jigsaw, and Wine (Shi et al., 2021) subjected to simulated distribution shifts induced by Gaussian noise, uniform noise, and randomly perturbed feature values(Wang et al., 2020).
Dataset Splits No The paper describes the overall TTT paradigm which includes source domain training and target domain adaptation, but does not provide specific train/validation/test split percentages or counts for the datasets used in their experiments. It references benchmark datasets, but not how they specifically partitioned them into validation sets for their work.
Hardware Specification Yes Our experiments were conducted on a Linux server equipped with an A100 GPU.
Software Dependencies No The paper mentions that "Our code uses existing open-source implementations of existing methods" and outlines optimization rules like SGD, but it does not specify software names with version numbers (e.g., PyTorch 1.9, Python 3.8).
Experiment Setup Yes Unless otherwise specified, the default batch size used was 64. The number of logical neural network layers for learning logical connectives was set to 3 based on exploratory runs that tried values from 1 to 6. The number of conjunction and disjunction modules in each rule learning layer was set to 16. Temperature parameter in Equation 9 was set as 0.1.