A hyperlink relaxes the restriction in traditional link prediction that two nodes form a link. Instead, it allows an arbitrary number of nodes to jointly form a multiway relation. Hyperlink prediction is much harder than link prediction, as the total number of hyperlinks is O(2^n) instead of O(n^2).
This toolbox contains the source code of our paper:
M. Zhang, Z. Cui, S. Jiang, and Y. Chen, Beyond Link Prediction: Predicting Hyperlinks in Adjacency Space, Proc. AAAI Conference on Artificial Intelligence (AAAI-18). [Preprint]
which studies the hyperlink prediction problem. We proposed a novel Coordinated Matrix Minimization (CMM) algorithm, which predicts hyperlinks in the vertex adjacency space and achieves state-of-the-art performance on two tasks: recipe prediction and metabolic reaction prediction. This toolbox also includes implementations of various hyperlink prediction baselines.
Have a try by typing "Main_meta" in MATLAB!
The file "Main_meta.m" is the main program for running the metabolic reaction prediction experiments. The file "Main_dish" is for dish prediction. Please change experimental settings inside these two files.
The folder "data/" contains the preprocessed data used in the experiments. They are: iAB_RBC_283.mat, iAF1260b.mat, iAF692.mat, iHN637.mat, iIT341.mat and iJO1366.mat (each is a metabolic network model), and chuancai.txt, yuecai.txt (collections of Sichuan and Cantonese recipes). The folder "utils/" contains some preprocessing scripts, which may be useful if you want to experiment on more datasets.
By default, "Main_meta.m" will save results in "result/". There is a "Meta_Plot.m" in "utils/" which plots figures according to the results.
The software symnmf-master is needed and already included in "software/".
The software liblinear is required for some baselines. We have included a "liblinear.tar.gz" in "software/". Type:
tar -xvzf liblinear.tar.gz
to unzip it. Then please follow its README to install it (and its MATLAB version, see README in /matlab/). After installation, you need to change the names of "train.mexa64" and "predict.mexa64" to "liblinear_train.mexa64" and "liblinear_predict.mexa64" respectively, otherwise they will cause conflictions.
The software libFM is required for some baselines. We have included a "libfm-1.42.src.tar.gz" in "software/". Type:
tar -xvzf libfm-1.42.src.tar.gz
to unzip it. Then, compile it by
cd libfm-1.42-src
make all
cp bin/libFM ../../../data/FM_temp/
This will compile and copy the executable libFM into "data/FM_temp".
If you find the code useful, please cite our paper:
@inproceedings{zhang2018beyond,
title={Beyond Link Prediction: Predicting Hyperlinks in Adjacency Space.},
author={Zhang, Muhan and Cui, Zhicheng and Jiang, Shali and Chen, Yixin},
booktitle={AAAI},
year={2018}
}
Muhan Zhang, muhan@wustl.edu 12/2/2017