Explaining Naive Bayes and Other Linear Classifiers with Polynomial Time and Delay

Authors: Joao Marques-Silva, Thomas Gerspacher, Martin Cooper, Alexey Ignatiev, Nina Narodytska

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results demonstrate the performance gains of the new algorithms when compared with earlier work. The experimental results also investigate ways to measure the quality of heuristic explanations.
Researcher Affiliation Collaboration Joao Marques-Silva1, Thomas Gerspacher2, Martin C. Cooper1 1IRIT, CNRS, University of Toulouse III, France 2ANITI, University of Toulouse, France {joao.marques-silva,thomas.gerspacher,cooper}@irit.fr Alexey Ignatiev Monash University, Australia alexey.ignatiev@monash.edu Nina Narodytska VMware Research, CA, USA nnarodytska@vmware.com
Pseudocode Yes Algorithm 1: Finding one explanation; Algorithm 2: Finding all explanations; Algorithm 3: Entering a valid state
Open Source Code Yes The source code of XPXLC as well as the datasets, a demo and accompanying documentation are available at https://github.com/jpmarquessilva/expxlc.
Open Datasets Yes We selected a set of widely-used, publicly available, datasets from [37, 28, 13]. The total number of datasets used is 37. (All the datasets and the trained classifiers are available in the online repository.)
Dataset Splits No For each dataset, we trained a Naive Bayes classifier13 using 80% of the training data. The average test accuracy assessed for the 20% remaining instances is 77.7%. No explicit mention of a validation split.
Hardware Specification Yes XPXLC was tested in Debian Linux on an Intel Xeon CPU 5160 3.00 GHz with 64 GByte of memory.
Software Dependencies No The paper mentions 'Debian Linux' as the operating system and 'scikit-learn [33]' for the Naive Bayes classifier, but does not specify exact version numbers for these or any other software dependencies.
Experiment Setup No The paper describes using a Naive Bayes classifier trained on 80% of the data, but does not provide specific experimental setup details such as hyperparameters (e.g., learning rate, batch size, optimizer settings) or model initialization.