This repository is about steps of this semi-supervised learning algorithm.

• Original paper: MixMatch - A Holistic Approach to Semi-Supervised Learning by David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver and Colin Raffel.
• Code provided by Google Research is here.

This algorithm is originally designed for image classification and usually requires CUDA support. Together with Yang Wan and Minchenxi Zhou, I used this approach and modified the code for tabular data and a CPU-only environment in my internship. I cannot upload code here because of confidentiality requirement, but I can share my understanding of this algorithm because it's public content:)

• Categorized as labeled data & unlabeled data
• Data preprocessing
  • Drop features where over 95% of the data are missing
  • Impute missing values: -1 for int (discrete values), mean for float (continuous values)
  • Construct matrices after dimension seperation
  • The final data should be in a dimension of 4
    • sample size
    • RGB parameter (3 if colored, 1 if grey scale or non-image data)
    • matrix length
    • matrix width

• Augmentation times: 1 for labeled data; K (hyper parameter) for unlabeled data
• For image data
  • Strong augmentation: sharpening, adjusting saturation, and adjusting color temperature
  • Weak augmentation: translation, rotation, and cropping
• For tabular data
  • Random flipping and cropping of matrices (substituting the margins with all 0)

• Generate pseudo labels for unlabeled data with models. Try multiple times and take the mean as the final result
• For image data, Wide-ResNet-28 (28 layers of wide residual networks) is commonly used. However, technically other unsupervised learning models could work too, depending on the data format
  • Wide-Res-Net structure
  

• Minimize entropy and transform the predictions to a one-hot distribution. Pick the one with noticeably largest value

• Put together X and U as W; then shuffle W
  • X：augmented labeled dataset and their labels
  • U：augmented unlabeled dataset and their pseudo labels
  • hyper-parameter α = 0.75
    • λ ~ Beta(α, α)
    • λ' = max(λ, 1-λ)
    • x' = λ'x1 + (1-λ')x2
    • p' = λ'p1 + (1-λ')p2

• Mix up every element in W
• Take maximums for λ & 1-λ, so that (x1, p1) still takes up a principal component in the final result after mixup. In this way, X's can still represent labeled data while U's can represent unlabeled ones

• Calculate cross entropy for labeled data
• Calculate the L2 distance between predictions and pseudo labels for unlabeled data
• L = Lx + λuLu
  • hyper-parameter λu = 100

• Train the model with Wide-Res-Net again
• Tune parameters until the model performs well