Code for "Hinge-loss Markov Random Fields: Convex Inference for Structured Prediction." Stephen H. Bach, Bert Huang, Ben London, and Lise Getoor. Uncertainty in Artificial Intelligence (UAI) 2013
Groovy Java Matlab Python Shell
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
data
src/main
.gitignore
README.md
pom.xml
runCollectiveClassification.sh
runImageReconstruction.sh
runPreferencePrediction.sh
runSocialTrustPrediction.sh
runTiming.sh

README.md

bach-uai13-code

Code for "Hinge-loss Markov Random Fields: Convex Inference for Structured Prediction." Stephen H. Bach, Bert Huang, Ben London, and Lise Getoor. Uncertainty in Artificial Intelligence (UAI) 2013

Please cite this work as

@inproceedings{bach:uai13,
	author = "Bach, Stephen H. and Huang, Bert and London, Ben and Getoor, Lise",
	title = "Hinge-loss {M}arkov Random Fields: Convex Inference for Structured Prediction",
	booktitle = "{Uncertainty in Artificial Intelligence (UAI)}",
	year = "2013"
}

Instructions

Prerequisites

This software depends on Java 6 and Maven 3 (http://maven.apache.org). Matlab and Python (>=2.7) are also required to process the results.

To run Bayesian probabilistic matrix factorization for the preference-prediction experiment, you must download and compile the Bayesian probabilistic tensor factorization package (http://www.cs.cmu.edu/~lxiong/bptf/bptf.html) and place the export_bptf directory in src/main/matlab.

PSL Library

The algorithms for these experiments are implemented in the PSL library, version 1.1 (https://github.com/linqs/psl/tree/1.1). It will be downloaded by Maven automatically.

Executing

Each experiment from the paper can be run from one of the shell scripts in the root directory.

Data

All data is in the data directory. The four data sets are described and credited below.

Citeseer

This data set contains a selection of the CiteSeer data set (http://citeseer.ist.psu.edu/).

These papers are classified into one of the following six classes:

Agents
AI
DB
IR
ML
HCI

The papers were selected in a way such that in the final corpus every paper cites or is cited by at least one other paper. There are 3312 papers in the whole corpus.

After stemming and removing stopwords we were left with a vocabulary of size 3703 unique words. All words with document frequency less than 10 were removed.

The citeseer.content file contains descriptions of the papers in the following format:

<paper_id> <word_attributes>+ <class_label>

The first entry in each line contains the unique string ID of the paper followed by binary values indicating whether each word in the vocabulary is present (indicated by 1) or absent (indicated by 0) in the paper. Finally, the last entry in the line contains the class label of the paper.

The citeseer.cites file contains the citation graph of the corpus. Each line describes a link in the following format:

<ID of cited paper> <ID of citing paper>

Each line contains two paper IDs. The first entry is the ID of the paper being cited and the second ID stands for the paper which contains the citation. The direction of the link is from right to left. If a line is represented by "paper1 paper2" then the link is "paper2->paper1".

parse.py parses citeseer.content and citeseer.cites to create additional data files. It is not necessary to run this again.

Please cite this data set as

@Article{sen:aimag08,
	author       = "Sen, Prithviraj and Namata, Galileo Mark and Bilgic, Mustafa and Getoor, Lise and Gallagher, Brian and Eliassi-Rad, Tina",
	title        = "Collective Classification in Network Data",
	journal      = "AI Magazine",
	number       = "3",
	volume       = "29",
	pages        = "93--106",
	year         = "2008"
}

Cora

This data set contains a selection of the Cora data set (http://www.research.whizbang.com/data).

The Cora dataset consists of Machine Learning papers. These papers are classified into one of the following seven classes:

Case_Based
Genetic_Algorithms
Neural_Networks
Probabilistic_Methods
Reinforcement_Learning
Rule_Learning
Theory

The papers were selected in a way such that in the final corpus every paper cites or is cited by atleast one other paper. There are 2708 papers in the whole corpus.

After stemming and removing stopwords we were left with a vocabulary of size 1433 unique words. All words with document frequency less than 10 were removed.

The cora.content file contains descriptions of the papers in the following format:

<paper_id> <word_attributes>+ <class_label>

The first entry in each line contains the unique string ID of the paper followed by binary values indicating whether each word in the vocabulary is present (indicated by 1) or absent (indicated by 0) in the paper. Finally, the last entry in the line contains the class label of the paper.

The cora.cites file contains the citation graph of the corpus. Each line describes a link in the following format:

<ID of cited paper> <ID of citing paper>

Each line contains two paper IDs. The first entry is the ID of the paper being cited and the second ID stands for the paper which contains the citation. The direction of the link is from right to left. If a line is represented by "paper1 paper2" then the link is "paper2->paper1".

parse.py parses cora.content and cora.cites to create additional data files. It is not necessary to run this again.

Please cite this data set as

@Article{sen:aimag08,
	author       = "Sen, Prithviraj and Namata, Galileo Mark and Bilgic, Mustafa and Getoor, Lise and Gallagher, Brian and Eliassi-Rad, Tina",
	title        = "Collective Classification in Network Data",
	journal      = "AI Magazine",
	number       = "3",
	volume       = "29",
	pages        = "93--106",
	year         = "2008"
}

Epinions

This data set is a snowball sample from the Epinions trust data set (http://snap.stanford.edu/data/soc-sign-epinions.html).

From the website: "This is who-trust-whom online social network of a a general consumer review site Epinions.com. Members of the site can decide whether to ''trust'' each other. All the trust relationships interact and form the Web of Trust which is then combined with review ratings to determine which reviews are shown to the user."

Please cite the original data set as

@inproceedings{leskovec:chi10,
	author = "Leskovec, Jure and Huttenlocher, Daniel and Kleinberg, Jon},
	booktitle = "{28th ACM Conference on Human Factors in Computing Systems (CHI)}",
	pages = "1361--1370",
	title = "Signed Networks in Social Media",
	year = "2010"
}

Jester

Data set of ratings of jokes (http://goldberg.berkeley.edu/jester-data/).

The original data (data/jester/ratings/jester-1.csv and data/jester/joketext/joketext.txt) are parsed into the 50/50 test/train splits with 50% ratings observed and unobserved using data/jester/ratings/parse.py and data/jester/joketext/computeSim.py. It is not necessary to re-run these scripts. The users to model during training and testing are listed in data/jester/users-te-1000.txt and data/jester/users-tr-1000.txt.

Please cite the original data set as

@article{goldberg:ir01,
	author = "Goldberg, Ken and Roeder, Theresa and Gupta, Dhruv and Perkins, Chris",
	journal = "Information Retrieval",
	number = "2",
	pages = "133--151",
	title = "Eigentaste: A Constant Time Collaborative Filtering Algorithm",
	volume = "4",
	year = "2001"
}

Olivetti

Faces from the AT&T database.

Images are 64x64 and have been preprocessed to a vector format (1 face per row) and pixel intensities have been normalized from 0-255 to 0-1.

Please cite this data set as

@inproceedings{samaria:acv94,
	author = "Samaria, F. S. and Harter, A. C.",
	booktitle = "{Proceedings of 1994 IEEE Workshop on Applications of Computer Vision}",
	title = "Parameterisation of a stochastic model for human face identification",
	year = "1994"
}

Caltech-101 Faces

Faces from the Caltech-101 image data set (http://www.vision.caltech.edu/Image_Datasets/Caltech101/).

Images are the center 64x64 pixels of the original face images. They have been preprocessed to a vector format (1 face per row) and pixel intensities have been normalized from 0-255 to 0-1.

Please cite this data set as

@article{feifei:cviu07,
	author = "Fei-Fei, Li and Fergus, Rob and Perona, Pietro",
	journal = "Comput. Vis. Image Underst.",
	number = "1",
	pages = "59--70",
	title = "Learning generative visual models from few training examples: An incremental {B}ayesian approach tested on 101 object categories",
	volume = "106",
	year = "2007"
}