Exchangeability-Aware Sum-Product Networks

Authors: Stefan Lüdtke, Christian Bartelt, Heiner Stuckenschmidt

IJCAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We introduce a structure learning algorithm for XSPNs and empirically show that they can be more accurate than conventional SPNs when the data contains repeated, interchangeable parts. and 4 Experimental Evaluation
Researcher Affiliation Academia Stefan L udtke1 , Christian Bartelt1 and Heiner Stuckenschmidt2 1Institute for Enterprise Systems, University of Mannheim, Germany 2Data and Web Science Group, University of Mannheim, Germany
Pseudocode Yes Algorithm 1 Learn XSPN
Open Source Code Yes Our implementation2 of XSPNs is based on the SPFlow library [Molina et al., 2019] for Python. 2Available at https://github.com/stefanluedtke/XSPNFlow
Open Datasets Yes The senate dataset contains all 720 roll call votes in the Senate of the 116th United States Congress, taken from [Lewis et al., 2021]. and 17th German federal parliament (bundestag dataset, data taken from [Bergmann et al., 2018]).
Dataset Splits Yes We performed an exhaustive grid search of the SPN hyperparameters. and Table 1 shows the experimental results. For the SPN and XSPN, results that are significantly better than the other (X)SPN model are printed in bold (paired t-test, p < 0.05).
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory, or specific cloud instance types) used for running the experiments.
Software Dependencies No Our implementation of XSPNs is based on the SPFlow library [Molina et al., 2019] for Python. The paper mentions software tools like SPFlow and Python but does not provide specific version numbers for them (e.g., Python 3.x, SPFlow vX.Y).
Experiment Setup Yes We performed an exhaustive grid search of the SPN hyperparameters. Specifically, we varied the g-test threshold values ρ {5, 15}, the minimum number of instances m {20, 200}, and the significance level of the χ2-test for exchangeability p {0.05, 0.1, 0.2, 0.4}. For all models, we used Laplace smoothing with α = 0.1.