Explainable Data Decompositions with Desc

This repository provides a Julia library that implements the Desc algorithm. By leveraging the Bayesian information criterion and maximum entropy modeling, Desc efficiently discovers concise sets of informative higher-order feature interactions (i.e., patterns). Desc highlights commonalities and differences between groups by associating patterns with sets of groups if the pattern is characteristic for the groups.

The code is a from-scratch implementation of algorithms described in the paper.

Dalleiger, S. and Vreeken, J. 2020. Explainable Data Decompositions. Proceedings of the AAAI Conference on Artificial Intelligence, pp. 3709–3716. https://doi.org/10.1609/aaai.v34i04.5780

Please consider citing the paper.

Contributions are welcome.

Installation

To install the library from the REPL:

julia> using Pkg; Pkg.add(url="https://github.com/sdall/desc.git")

To install the library from the command line:

julia -e 'using Pkg; Pkg.add(url="https://github.com/sdall/desc.git")'

To set up the command line interface (CLI) located in bin/desc.jl:

Clone the repository:

git clone https://github.com/sdall/desc

Install the required dependencies including the library:

julia -e 'using Pkg; Pkg.add(path="./desc"); Pkg.add.(["Comonicon", "CSV", "GZip", "JSON"])'

Usage

A typical usage of the library is:

julia> using Desc: desc, patterns
julia> p = desc(X, y)
julia> patterns.(p)

For more information, please see the provided documentation:

help?> desc

A typical usage of the command line interface is:

chmod +x bin/desc.jl
bin/desc.jl dataset.dat.gz dataset.labels.gz > output.json

The output contains patterns and executiontime in seconds (cf. --measure-time for details). For more information regarding usage, additional options, or input format, please see the provided documentation:

bin/desc.jl --help