Quadratization and Roof Duality of Markov Logic Networks
Authors: Roderick Sebastiaan de Nijs, Christian Landsiedel, Dirk Wollherr, Martin Buss
JAIR 2016 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental evaluation of the quadratization techniques in combination with the QPBO algorithm show that large benefits in performance can be attained on real-world problems, and that the more sophisticated quadratization techniques deliver better results than the one employed by Fierens et al. (2013). The combination of quadratization and QPBO is shown to be a competitive strategy for approaching the MAP problem in MLNs. (...) In Section 5, a thorough computational evaluation of the approach is performed on datasets from literature as well as new problems. (...) We evaluate our approach on various standard MLNs and datasets as well as additional problems. |
| Researcher Affiliation | Academia | Roderick de Nijs EMAIL Institute of Computer Science Albrechtstr. 28, 49076 Osnabrück, Germany; Christian Landsiedel EMAIL Dirk Wollher EMAIL Martin Buss EMAIL Lehrstuhl f ur Steuerungsund Regelungstechnik Theresienstr. 90, 80333 München, Germany |
| Pseudocode | No | The paper describes various algorithms and transformations using mathematical notation and textual descriptions (e.g., Section 4: Quadratization), but it does not include any clearly labeled pseudocode or algorithm blocks with structured steps. |
| Open Source Code | No | The paper does not contain any explicit statements about releasing source code, nor does it provide links to a code repository. The text does not mention code in supplementary materials or upon request. |
| Open Datasets | Yes | We evaluate our approach on various standard MLNs and datasets as well as additional problems. The characteristics of these problems are summarized in Table 1. A first set of datasets is similar to the ones employed in the evaluation of state-of-the-art engines Tuffy (Niu et al., 2011) and Rock It (Noessner et al., 2013). The link prediction problem on the UWCSE dataset (LP) tries to find relations between faculty members and students. The relational classification (RC) on the Cora dataset determines the category of research papers. The information extraction (IE) problem models how to obtain dataset records from parsed sources. The web KB dataset is used to predict to which university department a website belongs, given its hyperlink relations and contained words (KB). The entity resolution (ER) problem on the Cora dataset is obtained from the Alchemy website. The goal of this problem is to identify citations referring to the same paper. Because no trained model is available for this problem, it is trained with Alchemy using the first of the five available splits for evaluation (Singla & Domingos, 2006a). |
| Dataset Splits | Yes | The entity resolution (ER) problem on the Cora dataset is obtained from the Alchemy website. The goal of this problem is to identify citations referring to the same paper. Because no trained model is available for this problem, it is trained with Alchemy using the first of the five available splits for evaluation (Singla & Domingos, 2006a). (...) A 90 90 pixels random binary image is used as evidence, where each pixel has a 50% chance of being on or off. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU models, CPU types, or memory amounts used for running the experiments. It only discusses the software engines and their configurations. |
| Software Dependencies | No | We compare our approach with the MAP-inference solvers Alchemy, Tuffy and Rock It. Alchemy is the original solver for MLNs (...) The Tuffy system (Niu, R e, Doan, & Shavlik, 2011) (...) The Rock It system (Noessner, Niepert, & Stuckenschmidt, 2013). While these tools are named, specific version numbers are not provided. |
| Experiment Setup | Yes | Alchemy and Tuffy were run for an increasing number of flips until no significant advances were made. Rock It was run with relative gaps 1 10 n, n = 9, 8, . . . until convergence is achieved within an hour. (...) The unary rules are given weight 1.0. To ensure that the terms of the smoothing rules do not cancel out, a rule for the on pixels is given a weight of 0.35 and the rule for the off pixels 0.3. (...) Using the ASM and GRD quadratization for the higher-order problems, using improve on the residual problem until no advances were made for 20 iterations. |