Abduction-Based Explanations for Machine Learning Models

Authors: Alexey Ignatiev, Nina Narodytska, Joao Marques-Silva1511-1519

AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The experimental results, obtained on well-known datasets, validate the scalability of the proposed approach as well as the quality of the computed solutions.
Researcher Affiliation Collaboration Alexey Ignatiev,1,3 Nina Narodytska,2 Joao Marques-Silva1 1Faculty of Science, University of Lisbon, Portugal 2VMware Research, CA, USA 3ISDCT SB RAS, Irkutsk, Russia {aignatiev,jpms}@ciencias.ulisboa.pt, nnarodytska@vmware.com
Pseudocode Yes Algorithm 1: Computing a subset-minimal explanation; Algorithm 2: Computing a smallest size explanation
Open Source Code No The paper mentions that its prototype implementation is "written in Python" and uses other open-source tools (e.g., Py SMT, RC2), but it does not provide an explicit statement or a link to its own source code for the methodology described in the paper.
Open Datasets Yes The benchmarks considered include the well-known text-based datasets from the UCI Machine Learning Repository1 and Penn Machine Learning Benchmarks2, as well as the widely used MNIST digits database3. ... 1https://archive.ics.uci.edu/ml/ 2https://github.com/EpistasisLab/penn-ml-benchmarks/ 3http://yann.lecun.com/exdb/mnist/
Dataset Splits No The paper states that neural networks were trained on datasets and evaluated on samples, but it does not explicitly specify the training/validation/test split percentages, absolute sample counts for each split, or detailed splitting methodology needed for reproduction.
Hardware Specification Yes all benchmarks were ran on a Macbook Pro having an Intel Core i7 2.8GHz processor with 8GByte of memory on board.
Software Dependencies Yes CPLEX 12.8.0 (IBM ILOG 2018) is used as a MILP oracle accessed via its official Python API.
Experiment Setup Yes Time limit was set to 1800 seconds while memory limit was set to 4GByte. ... Each neural network considered has one hidden layer with i {10, 15, 20} neurons.