Disentanglement: Confounder-Agonistic (POT-AID/Contrastive) Baselines

Code for domain-invariant and confounder-invariant feature learning.

• Works for any confounder ("batch", "scanner", "site", "modality" etc.): just make sure your meta CSVs contain both label (the task) and confounder columns.
• Includes:
  • POT-AID (Partial Orthogonalization + Adversarial Invariant Disentangling)
  • Contrastive Disentanglement (HSIC + Supervised Contrastive)

Data organization

• Place all .npy and meta .csv files in a directory (features/ by default).
• Each meta CSV must have columns:
  • label: class/task label
  • confounder: the variable you want to remove/invariantize