Tabular TTA OOD

This repository contains the implementation of an Advanced Machine Learning course project on out-of-distribution (OOD) generalization for tabular data.

The project studies test-time adaptation (TTA) on tabular benchmarks derived from TableShift, and compares four methods under a unified local CSV-based experimental pipeline:

• ERM: source-only empirical risk minimization
• Tent: test-time entropy minimization
• Recon-TTA: test-time masked feature reconstruction
• DARTA: a dual-objective test-time adaptation method that combines entropy minimization and masked feature reconstruction

The codebase is designed for reproducible experimentation in a Linux server environment and uses the course-provided offline TableShift dataset package rather than the online TableShift loader.

---

1. Project Overview

The goal of this project is to study whether test-time adaptation can improve OOD performance on tabular tasks.

Unlike image data, tabular data typically involve:

• heterogeneous feature spaces
• fixed feature semantics
• strong class imbalance
• complex domain shifts
• large differences in dataset scale

These characteristics make OOD generalization on tabular data both practically important and technically challenging.

This project builds a unified pipeline to train source-domain models, adapt them at test time, and compare their performance on ID and OOD splits.

---

2. Datasets

This project uses the 6 required datasets from the course-provided TableShift package:

• assistments
• nhanes_lead
• brfss_diabetes
• acsfoodstamps
• physionet
• acsunemployment

Each dataset is converted into the following local format:

data/processed/<dataset_name>/
├── train.csv
├── val.csv
├── test_id.csv
└── test_ood.csv