Ecosyste.ms: Awesome

An open API service indexing awesome lists of open source software.

Awesome Lists | Featured Topics | Projects

https://github.com/guicho271828/remlic

Reimplementation of an interpretable machine learning system MLIC in Common Lisp
https://github.com/guicho271828/remlic

Last synced: 29 days ago
JSON representation

Reimplementation of an interpretable machine learning system MLIC in Common Lisp

Awesome Lists containing this project

README 

        
# reMLIC



Reimplementation of MLIC [1], an interpretable machine learning system in Common Lisp.



[1] Malioutov, Dmitry, and Kuldeep S. Meel. "Mlic: A maxsat-based framework for learning interpretable classification rules." International Conference on Principles and Practice of Constraint Programming. Springer, Cham, 2018.



## Installation



``` shell

ros install numcl/constantfold numcl/specialized-function numcl/gtype numcl/numcl guicho271828/remlic

```



## Usage



    mlic           input output rules &rest args &key accuracy sparsity diversity unlabeled oneshot verbose evaluate val-input val-output



Runs a MLIC algorithm on a bit array input and a bit-array output, both in [NUMCL-compatible format](https://github.com/numcl/numcl/blob/master/doc/DETAILS.org#representation).

The first dimension is treated as the batch dimension.



`accuracy` and `sparsity` controls the penalization weights in MLIC that

balances the accuracy and the interpretability.



MLIC input is, similar to a fully-connected NN, not structured. Therefore,

arrays with rank larger than 2 are internally reshaped into an 2D array.

MLIC also runs the training for each output dimension separately.



    mlic-core      input output rules &rest regularizers class evaluate verbose timelimit val-input val-output solver-args



This is the base function called by mlic / mlic-iterative. All keywords accepted

by this function is also accepted by mlic / mlic-iterative.



`solver-args` is lambda-list arguments to `cl-maxsat:solve`, which can specify the solver.

The default value is `(:maxsat-competition :year 2017 :track :complete :name :maxhs)`.



`class` specifies how we model the constraints. `cnf` is the standard mode used in the original MLIC paper.

However, we also add several other classes that may better characterize your dataset: `cnf`, `dnf`, `eqv`, and so on.

See the corresponding documentation string in `src/2constraints.lisp`.



    mlic-iterative input output rules &rest args &key initial-samples max-samples callback by timelimit method evaluate val-input val-output



Since MLIC is computationally intensive (using Max-SAT solver), we made an

iterative variant that starts form the small dataset and increase the sample

size until the validation accuracy meets a certain threshold.

`callback` specifies the function that specifies the condition.

For details, see the implementation of function `validation-above-90`.



## Dependencies

This library is at least tested on implementation listed below:



+ SBCL 1.4.12 on X86-64 Linux 4.4.0-141-generic (author's environment)



Also, it depends on the following libraries:



+ cl-maxsat by *Masataro Asai* : Common Interface to MAX-SAT Solvers from Common Lisp

+ numcl by *Masataro Asai* : Numpy clone in Common Lisp

+ trivial-timeout by *Gary Warren King* : OS and Implementation independent access to timeouts

+ trivia by *Masataro Asai* :

    NON-optimized pattern matcher compatible with OPTIMA, with extensible optimizer interface and clean codebase

+ alexandria by *Nikodemus Siivola , and others.* :

    Alexandria is a collection of portable public domain utilities.

+ iterate by ** :

    Jonathan Amsterdam's iterator/gatherer/accumulator facility



## Author, License, Copyright



Licensed under LGPL v3.



Copyright (c) 2019 Masataro Asai ([email protected])

Copyright (c) 2019 IBM Corporation