Self-supervised Learning for Tabular Data

• NIPS (2020) VIME: Extending the Success of Self- and Semi-supervised Learning to Tabular Domain

  

  

• ICLR (2021) Self-Supervision Enhanced Feature Selection with Correlated Gates

  • Most structure is similar with the above one, one difference is on the distribution of the mask vector (using feature correlation information)

    

  • The correlated gating procedure in the proposed method prevents the pretext tasks from being solved by only exploiting trivial relationships among features by increasing the probability that highly-correlated features are masked simultaneously.

• CIKM (2022) Local Contrastive Feature Learning for Tabular Data

  • Our model proposes a new feature reordering using feature-feature correlations and the order of features is determined by a depth-first search
  • Then we applies local feature learning to reordered feature subsets by using CNN + contrastive learning.
  • We apply a data augmentation on sample subsets through masking which randomly generates a binary matrix with a batch of data related to input feature subsets

• ICLR (2022) SCARF: Self-Supervised Contrastive Learning using Random Feature Corruption