tf-fsvd

TensorFlow Implementation of Functional Singular Value Decomposition for paper Fast Graph Learning with Unique Optimal Solutions

Cite

If you find our code useful, you may cite us as:

@inproceedings{haija2021fsvd,
  title={Fast Graph Learning with Unique Optimal Solutions},
  author={Sami Abu-El-Haija AND Valentino Crespi AND Greg Ver Steeg AND Aram Galstyan},
  year={2021},
  booktitle={arxiv:2102.08530},
}

Introduction

This codebase contains TensorFlow implementation of Functional SVD, an SVD routine that accepts objects with 3 attributes: dot, T, and shape. The object must be able to exactly multiply an (implicit) matrix M by any other matrix. Specifically, it should implement:

1. dot(M1): should return M @ M1
2. T: property should return another object that (implicitly) contains transpose of M.
3. shape: property should return the shape of the (implicit) matrix M.

In most practical cases, M is implicit i.e. need not to be exactly computed. For consistency, such objects could inherit the abstract class ProductFn.

Simple Usage Example

Suppose you have an explicit sparse matrix mat

import scipy.sparse
import tf_fsvd

m = scipy.sparse.csr_mat( ... )
fn = tf_fsvd.SparseMatrixPF(m)

u, s, v = tf_fsvd.fsvd(fn, k=20)  # Rank 20 decomposition

The intent of this utility is for implicit matrices. For which, you may implement your own ProductFn class. You can take a look at BlockWisePF or WYSDeepWalkPF.

File Structure / Documentation

• File tf_fsvd.py contains the main logic for TensorFlow implementation of Functional SVD (function fsvd), as well as a few classes for constructing implicit matrices.
  • SparseMatrixPF: when implicit matrix is a pre-computed sparse matrix. Using this class, you can now enjoy the equivalent of tf.linalg.svd on sparse tensors :-).
  • BlockWisePF: when implicit matrix is is column-wise concatenation of other implicit matrices. The concatenation is computed by suppling a list of ProductFn's
• Directory implementations: contains implementations of simple methods employing fsvd.
• Directory baselines: source code adapting competitive methods to produce metrics we report in our paper (time and accuracy).
• Directory experiments: Shell scripts for running baselines and our implementations.
• Directory results: Output directory containing results.

Running Experiments

ROC-AUC Link Prediction over AsymProj/WYS datasets

The AsymProj datasets are located in directory datasets/asymproj.

You can run the script for training on AsympProj datasets and measuring test ROC-AUC as:

python3 implementations/linkpred_asymproj.py

You can append flag --help to above to see which flags you can set for changing the dataset or the SVD rank.

You can run sweep on svd rank, for each of those datasets, by invoking:

# Sweep fSVD rank (k) on 4 link pred datasets. Make 3 runs per (dataset, k)
# Time is dominated by statement `import tensorflow as tf`
python3 experiments/fsvd_linkpred_k_sweep.py | bash  # You may remove "| bash" if you want to hand-pick commands.

# Summarize results onto CSV
python3 experiments/summarize_svdf_linkpred_sweep.py > results/linkpred_d_sweep/fsvd.csv

# Plot the sweep curve
python3 experiments/plot_sweep_k_linkpred.py

and running all printed commands. Alternatively, you can pipe the output of above to bash. This should populate directory results/linkpred_d_sweep/fsvd/.

Baselines

• You can run the Watch Your Step baseline as:

   bash experiments/baselines/run_wys.sh
  

  which runs only once for every link prediction dataset. Watch Your Step spends some time computing the transition matrix powers (T^2, .., T^5).

• You can run NetMF baselines (both approximate and exact) as:

  bash experiments/baselines/run_netmf.sh
  

• You can run node2vec baseline as:

  experiments/baselines/run_n2v.sh
  

Classification Experiments over Planetoid Citation datasets

These datasets are from the planetoid paper. To obtain them, you should clone their repo:

mkdir -p ~/data
cd ~/data
git clone git@github.com:kimiyoung/planetoid.git

You can run the script for training and testing on planetoid datasets as:

python3 implementations/node_ssc_planetoid.py

You can append flag --help to above to see which flags you can set for changing the dataset or the number of layers.

You can sweep the number of layers running:

# Directly invokes python many times
LAYERS=`python3 -c "print(','.join(map(str, range(17))))"`
python3 experiments/planetoid_hp_search.py --wys_windows=1 --wys_neg_coefs=1 --layers=${LAYERS}

The script experiments/planetoid_hp_search.py directly invokes implementations/node_ssc_planetoid.py. You can visualize the accuracy VS depth curve by running:

python3 experiments/plot_sweep_depth_planetoid.py

Link Prediction for measuring HITS@20 for Drug-Drug Interactions Network

You can run our method like:

python3 implementations/linkpred_ddi.py

This averages 10 runs (by default) and prints mean and standard deviation of validation and test metric (HITS@20)