Znn is an Elixir implementation of a nearest-neighbors text classifier that uses the Normalized Compression Distance. This approach is detailed in

Jiang et al., “Low-Resource” Text Classification: A Parameter-Free Classification Method with Compressors

Provide a training sample of labeled text data in the form

> training_set = [
    %{label: "clickbait", text: "21 Outrageously Creative Ways To Make Meatballs"},
    %{label: "non_clickbait", text: "H5N1 Avian Flu virus has mutated, study says"},
    ...
  ]

To predict the label of a new piece of text, call the predict function:

> text = "10 Amazing Ways To Grow Bigger Pumpkins"
> Znn.predict(text, training_set)
"clickbait"

See the included Livebooks in /livebooks for applications of Znn to public data sets.

• Precompute the compression sizes and store them with the label and text. One-the-fly compression is a significant fraction of the prediction run time.
• Allow other compression algorithms to be used, including no compression (identity function).
• Adaptively compute parallel chunk size through warm-up runs.
• Add multi-topic classification: handle a list of labels and include an appropriate voting scheme.