Exploiting Determinism to Scale Relational Inference
Authors: Mohamed Ibrahim, Christopher Pal, Gilles Pesant
AAAI 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on real-world applications show how PR substantially scales relational inference with a minor impact on accuracy. Finally we present experimental evaluations, followed by our conclusions. |
| Researcher Affiliation | Academia | Mohamed-Hamza Ibrahim and Christopher Pal and Gilles Pesant Department of Computer and Software Engineering, Ecole Polytechnique Montr eal 2500, Chemin de Polytechnique, Universite de Montr eal, Montr eal, Qu ebec, Canada |
| Pseudocode | Yes | Algorithm 1: PR-based Inference algorithm ˆ A and Algorithm 2: Combining PR with Lazy MC-SAT are provided. |
| Open Source Code | No | The paper mentions that they implemented PR as an extension to the Alchemy software but does not state that their specific implementation code is open-source or provide a link to it. For example: "All of the experiments were run on a cluster of nodes with 3.0 GHz Intel CPUs, 3 GB of RAM, RED HAT Linux 5.5, and we implemented PR as an extension to the Alchemy software (Kok et al. 2007)." |
| Open Datasets | Yes | Here we use the MLN model of (Davis and Domingos 2009) for a Yeast protein interaction problem. For the link prediction task, we used the MLN model available on the Alchemy website (excluding the 22 unit clauses) of the UW-CSE dataset from (Richardson and Domingos 2006). For the entity resolution experiment, we used the MLN model that is similar to the established one of Singla (Singla and Domingos 2006a) on the Cora dataset from (Poon, Domingos, and Sumner 2008). |
| Dataset Splits | Yes | In the training phase, we learned the weights using a preconditioned scaled conjugate gradient (PSCG) algorithm (Lowd and Domingos 2007) by performing a four-way cross-validation for protein interaction task, and a five-way cross-validation for both the link prediction and entity resolution tasks. |
| Hardware Specification | Yes | All of the experiments were run on a cluster of nodes with 3.0 GHz Intel CPUs, 3 GB of RAM, RED HAT Linux 5.5... |
| Software Dependencies | Yes | All of the experiments were run on a cluster of nodes with 3.0 GHz Intel CPUs, 3 GB of RAM, RED HAT Linux 5.5, and we implemented PR as an extension to the Alchemy software (Kok et al. 2007). |
| Experiment Setup | Yes | To obtain robust answers to the proposed questions, we did the following: 1) vary the number of objects in the domains, following methodology previously used (Poon, Domingos, and Sumner 2008); 2) use two pruning thresholds 0.001 and 0.01 for PR; and 3) vary the amounts of determinism in the models (i.e., starting with hard constraints that initially exist in MLN, we gradually increase the number of hard constraints and re-run the experiment). We ran each algorithm until it either converged or its number of iterations exceeded 10 000. |