Induction of Non-Monotonic Logic Programs to Explain Boosted Tree Models Using LIME

Authors: Farhad Shakerin, Gopal Gupta3052-3059

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments with UCI standard benchmarks suggest a significant improvement in terms of classification evaluation metrics.
Researcher Affiliation Academia Farhad Shakerin, Gopal Gupta Computer Science Department, The University of Texas at Dallas, Richardson, USA {farhad.shakerin,gupta}utdallas.edu
Pseudocode Yes Algorithm 1 Summarizing the FOIL algorithm; Algorithm 2 Linear Model Generation by LIME; Algorithm 3 FOLD Algorithm; Algorithm 4 Dataset Transformation with LIME
Open Source Code No The paper mentions that 'ALEPH v.5 has been ported into SWI-Prolog by (Riguzzi 2016)' with a GitHub link, but this is for a third-party tool (ALEPH) and not for the authors' own LIME-FOLD methodology. There is no concrete access provided for the LIME-FOLD source code.
Open Datasets Yes In this section, we present our experiments on UCI standard benchmarks (Lichman 2013). The ALEPH system (Srinivasan 2001) is used as the baseline. Lichman, M. 2013. UCI,ml repository, http://archive.ics. uci.edu/ml.
Dataset Splits Yes First, we run ALEPH on 10 different datasets using 5-fold crossvalidation setting. Second, each dataset is transformed as explained in Algorithm 4. Then the LIME-FOLD algorithm is run on a 5-fold cross-validated setting, and the classification metrics are reported.
Hardware Specification Yes All experiments were run on an Intel Core i7 CPU @ 2.7GHz with 16 GB RAM and a 64-bit Windows 10.
Software Dependencies Yes The FOLD algorithm is a Java application that uses JPL library to connect to SWI prolog. ALEPH v.5 has been ported into SWI-Prolog by (Riguzzi 2016).
Experiment Setup Yes We set ALEPH to use the heuristic enumeration strategy, and the maximum number of branch nodes to be explored in a branch-and-bound search to 500K. In this research we conducted all experiments using the Extreme Gradient Boosting (XGBoost) algorithm (Chen and Guestrin 2016).