Lifted Relational Neural Networks: Efficient Learning of Latent Relational Structures

Authors: Gustav Sourek, Vojtech Aschenbrenner, Filip Zelezny, Steven Schockaert, Ondrej Kuzelka

JAIR 2018 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we describe experiments performed on 78 datasets about organic molecules: the Mutagenesis dataset (Lodhi & Muggleton, 2005), four datasets from the predictive toxicology challenge, and 73 NCI datasets (Ralaivola, Swamidass, Saigo, & Baldi, 2005). We compare the performance of LRNNs with the state-of-the-art relational learners k FOIL (Landwehr, Passerini, De Raedt, & Frasconi, 2006) and n FOIL (Landwehr, Kersting, & Raedt, 2007), which respectively combine relational rule learning with support vector machines and with naive Bayes learning.
Researcher Affiliation Academia Gustav ˇSourek EMAIL Faculty of Electrical Engineering Czech Technical University in Prague Prague, Czech Republic Vojtˇech Aschenbrenner EMAIL Faculty of Mathematics and Physics Charles University Prague, Czech Republic Filip ˇZelezn y EMAIL Faculty of Electrical Engineering Czech Technical University in Prague Prague, Czech Republic Steven Schockaert EMAIL School of Computer Science & Informatics CardiffUniversity Cardiff, United Kingdom Ondˇrej Kuˇzelka EMAIL Department of Computer Science KU Leuven Leuven, Belgium
Pseudocode No The paper describes the weight learning algorithm textually in Section 3.4 'Weight Learning' but does not provide a clearly labeled pseudocode or algorithm block.
Open Source Code No The paper does not contain any explicit statements or links indicating that the source code for the described methodology is publicly available.
Open Datasets Yes In this section, we describe experiments performed on 78 datasets about organic molecules: the Mutagenesis dataset (Lodhi & Muggleton, 2005), four datasets from the predictive toxicology challenge, and 73 NCI datasets (Ralaivola, Swamidass, Saigo, & Baldi, 2005).
Dataset Splits Yes Figure 5: Prediction errors of LRNNs, k FOIL, n FOIL, MLN-boost and RDN-boost measured by cross-validation on 78 datasets about organic molecules.
Hardware Specification No The time for training an LRNN on a standard commodity machine with one CPU was in the order of a few hours for the larger NCI-GI datasets, and in the order of a few minutes for the smaller datasets such as Mutagenesis.
Software Dependencies No The paper mentions various methods and frameworks (e.g., backpropagation, stochastic gradient descent, MLNs, Problog, CILP++, k FOIL, n FOIL, MLN-boost, RDN-boost, Aleph), but does not provide specific version numbers for the software dependencies used in their implementation of LRNNs.
Experiment Setup Yes For all the reported experiments, we set the learning rate to 0.3 and training epochs to 3000.