Learning Relational Sum-Product Networks

Authors: Aniruddh Nath, Pedro Domingos

AAAI 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate the algorithm on three datasets; the RSPN learning algorithm outperforms Markov Logic Networks in both running time and predictive accuracy.
Researcher Affiliation Academia Aniruddh Nath and Pedro Domingos Department of Computer Science and Engineering University of Washington Seattle, WA 98195-2350, U.S.A. {nath, pedrod}@cs.washington.edu
Pseudocode Yes Algorithm 1 Learn RSPN(C, T, V )
Open Source Code No The paper does not provide an explicit statement about releasing the source code for the methodology described (Learn RSPN) nor does it provide a link to a code repository for their implementation.
Open Datasets Yes The UW-CSE database (Richardson and Domingos 2006), We generated artificial social networks in the Friends-and-Smokers domain (Singla and Domingos 2008), The test corpus consists of four short Python programming assignments from MIT ed X introductory programming course (6.00x) (Singh, Gulwani, and Solar-Lezama 2013)
Dataset Splits Yes We performed leave-one-out testing by area, testing on each area in turn using the model trained from the remaining four. The systems were trained on three programs and tested on the fourth
Hardware Specification No The paper does not provide specific hardware details (e.g., CPU/GPU models, memory) used for running the experiments.
Software Dependencies Yes For MLN inference, we used the MC-SAT algorithm, the default choice in ALCHEMY 2.0, with the default parameters.
Experiment Setup Yes To cluster instances in Learn RSPN, we used the EM implementation in SCIKIT-LEARN (Pedregosa et al. 2011), with two clusters. ... To discourage excessively finegrained decomposition during structure learning, we used a high threshold of 0.5 for the one-tailed p-value. For EDTs, we used the independent Bernoulli form, as described in example 1 in the main paper. All Bernoulli distributions were smoothed with a pseudocount of 0.1. For LSM, we used the example parameters in the implementation (Nwalks = 10, 000, π = 0.1; remaining parameters as specified by Kok and Domingos 2010).