Learning Ensembles of Cutset Networks
Authors: Tahrima Rahman, Vibhav Gogate
AAAI 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate, via a thorough experimental evaluation, that our new algorithms are superior to competing approaches in terms of test-set log-likelihood score and learning time. Second, we perform and report on a comprehensive empirical evaluation, comparing our new algorithms with several state-of-the-art systems such as sum-product networks with direct and indirect interactions (Rooshenas and Lowd 2014), latent tree models (Choi et al. 2011) and mixtures of cutset networks (Rahman, Kothalkar, and Gogate 2014) on a wide variety of benchmark datasets. |
| Researcher Affiliation | Academia | Tahrima Rahman and Vibhav Gogate Department of Computer Science The University of Texas at Dallas Richardson, TX 75080, USA. {tahrima.rahman,vibhav.gogate}@utdallas.edu |
| Pseudocode | Yes | Algorithm 1: CNet-Boosting |
| Open Source Code | No | The paper does not contain any explicit statement or link indicating that the source code for the described methodology is publicly available. |
| Open Datasets | Yes | We evaluated the performance of ECNets on 20 real world benchmark datasets (Lowd and Davis 2010; Gens and Domingos 2013; Rooshenas and Lowd 2014; Rahman, Kothalkar, and Gogate 2014; Van Haaren and Davis 2012; Davis and Domingos 2010) listed in Table 4. Table 4 explicitly lists dataset names, number of training examples, and provides citations to original sources, implying their public availability as standard benchmarks. |
| Dataset Splits | Yes | Table 4: Test set log-likelihood comparison. Dataset #Var #Train #Valid #Test ... NLTCS 16 16181 2157 3236 ... For the GBDE, we tried the following values for ϵ: {0.1,0.3,0.5,0.7,0.9} and βm: {0.01,0.05,0.2,0.4,0.6,0.8}. The best values were selected based on the accuracy attained on the validation set. In all three algorithms, we added a base model to the ensemble until the validation set likelihood decreased or the total number of boosting iterations reached 40. |
| Hardware Specification | Yes | We ran all our experiments on a quad-core Intel i7 2.7 GHz machine with 16GB RAM and ran each algorithm for 48 hours or until termination, whichever was earlier. |
| Software Dependencies | No | The paper does not specify any software dependencies with version numbers (e.g., programming languages, libraries, or specific tool versions). |
| Experiment Setup | Yes | For the GBDE, we tried the following values for ϵ: {0.1,0.3,0.5,0.7,0.9} and βm: {0.01,0.05,0.2,0.4,0.6,0.8}. In all three algorithms, we added a base model to the ensemble until the validation set likelihood decreased or the total number of boosting iterations reached 40. The maximum depth of CNets and MCNets was varied from 0 to 5 and the best depth was also chosen by the validation set. To learn simpler models than fully connected Chow Liu trees at the leaves of the base models (and thus avoid overfitting), we removed all edges from the Chow-Liu trees whose mutual information was smaller than 0.005. The parameters of the mixture were then optimized via the EM algorithm using the original training set for 50 iterations or until convergence. We varied the number of bags from 5 to 40 with increments of 5. The maximum depth of the OR trees was varied from 5 to 10. EM was run for 100 iterations or until convergence and no restarts were performed. |