Hinge-Loss Markov Random Fields and Probabilistic Soft Logic
Authors: Stephen H. Bach, Matthias Broecheler, Bert Huang, Lise Getoor
JMLR 2017 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | This paper brings together and expands work on scalable models for structured data that can be either discrete, continuous, or a mixture of both (Broecheler et al., 2010a; Bach et al., 2012, 2013, 2015b). The effectiveness of HL-MRFs and PSL has been demonstrated on many problems, including information extraction (Liu et al., 2016) and automatic knowledge base construction (Pujara et al., 2013), extracting and evaluating natural-language arguments on the Web (Samadi et al., 2016), high-level computer vision (London et al., 2013), drug discovery (Fakhraei et al., 2014) and predicting drug-drug interactions (Sridhar et al., 2016), natural language semantics (Beltagy et al., 2014; Sridhar et al., 2015; Deng and Wiebe, 2015; Ebrahimi et al., 2016), automobile-traffic modeling (Chen et al., 2014), recommender systems (Kouki et al., 2015), information retrieval (Alshukaili et al., 2016), and predicting attributes (Li et al., 2014) and trust (Huang et al., 2013; West et al., 2014) in social networks. ... We evaluate them on core relational learning and structured prediction tasks, such as collective classification and link prediction. We show that HL-MRFs offer predictive accuracy comparable to analogous discrete models while scaling much better to large data sets. ... In this section we evaluate the empirical performance of our MAP inference algorithm. ... To demonstrate the flexibility and effectiveness of learning with HL-MRFs, we test them on four diverse tasks: node labeling, link labeling, link prediction, and image completion. |
| Researcher Affiliation | Collaboration | Stephen H. Bach EMAIL Computer Science Department Stanford University Stanford, CA 94305, USA Matthias Broecheler EMAIL Data Stax Bert Huang EMAIL Computer Science Department Virginia Tech Blacksburg, VA 24061, USA Lise Getoor EMAIL Computer Science Department University of California, Santa Cruz Santa Cruz, CA 95064, USA |
| Pseudocode | Yes | Algorithm 1 MAP Inference for HL-MRFs Input: HL-MRF P(y|x), ρ > 0 Initialize y(L,j) as local copies of variables y(C,j) that are in φj, j = 1, . . . , m Initialize y(L,k+m) as local copies of variables y(C,k+m) that are in ck, k = 1, . . . , r Initialize Lagrange multipliers αˆi corresponding to copies y(L,ˆi), ˆi = 1, . . . , m + r while not converged do for j = 1, . . . , m do αj αj + ρ(y(L,j) y(C,j)) y(L,j) y(C,j) 1 ραj if ℓj(y(L,j), x) > 0 then y(L,j) arg miny(L,j) wj ℓj(y(L,j), x) pj + ρ 2 y(L,j) y(C,j) + 1 2 if ℓj(y(L,j), x) < 0 then y(L,j) Projℓj=0(y(C,j) 1 ραj) end if end if end for for k = 1, . . . , r do αk+m αk+m + ρ(y(L,k+m) y(C,k+m)) y(L,k+m) Projck(y(C,k+m) 1 ραk+m) end for for i = 1, . . . , n do yi 1 |copies(yi)| P yc copies(yi) yc + αc Clip yi to [0,1] end for |
| Open Source Code | Yes | An open source implementation, tutorials, and data sets are available at http://psl.linqs.org. ... Code is available at https://github.com/stephenbach/bach-jmlr17-code. ... Code is available at https://github.com/stephenbach/bach-jmlr17-code. ... An open source implementation, tutorials, and data sets are available at http://psl.linqs.org. |
| Open Datasets | Yes | We classify documents in citation networks using data from the Cora and Citeseer scientific paper repositories. ... We test this model on the Jester dataset, a repository of ratings from 24,983 users on a set of 100 jokes (Goldberg et al., 2001). ... We test these models using the experimental setup of Poon and Domingos (2011): we reconstruct images from the Olivetti face data set and the Caltech101 face category. |
| Dataset Splits | Yes | For each of 20 runs, we split the data sets 50/50 into training and testing partitions, and seed half of each set. ... We perform eight-fold cross-validation. ... We sample a random 2,000 users from the set of those who rated all 100 jokes, which we then split into 1,000 train and 1,000 test users. From each train and test matrix, we sample a random 50% to use as the observed features x; the remaining ratings are treated as the variables y. ... Following their experimental setup, we hold out the last fifty images and predict either the left half of the image or the bottom half. ... We train these HL-MRFs using SP with a 5.0 step size on the first 200 images of each data set and test on the last fifty. |
| Hardware Specification | Yes | All experiments are performed on a single machine with a 4-core 3.4 GHz Intel Core i7-3770 processor with 32GB of RAM. ... In our experiments, training each model takes about 45 minutes on a 12-core machine |
| Software Dependencies | Yes | We implement ADMM in Java and compare with the IPM in MOSEK (version 6) by encoding the entire MPE problem as a linear program or a second-order cone program as appropriate and passing the encoded problem via the Java native interface wrapper. |
| Experiment Setup | Yes | We weight local features with a parameter of 0.5 and choose parameters in [0, 1] for the relationship potentials representing a mix of more and less influential relationships. ... When training with SP and MPLE, we use 100 gradient steps and a step size of 1.0 (unless otherwise noted), and we average the iterates as in voted perceptron. For LME, we set C = 0.1. ... For prediction, we performed 2500 rounds of Gibbs sampling, 500 of which were discarded as burn-in. ... For inference with discrete MRFs, we perform 5000 rounds of Gibbs sampling, of which the first 500 are burn-in. ... we set the rank of the decomposition to 30 and use 100 iterations of burn in and 100 iterations of sampling. ... We train these HL-MRFs using SP with a 5.0 step size on the first 200 images of each data set and test on the last fifty. |