wangcunxiang / domain-representation-for-knowledge-graph-embedding

the project for Domain Representation for Knowledge Graph Embedding

DRE: Domain Representation in Knowledge Graph Embedding

Source code for DRE.

Data sets are WN18 and FB15K.

Usage

Compile the program

We use g++ to compile in the program.

g++ X.cpp -o X -O3 -pthread -march=native
train: X = ellipsoid
test: X = test_ellipsoid/test_R_ellipsoid/test_ST_ellipsoid

Run the program

####Train

The program combines training of TransE/TransR/STransE and DRE(TransE)/DRE(TransR)/DRE(STransE) in ellipsoid.cpp file. You can choose the baseline model from TransE/TransR/STransE, and the program will train the baseline model and the ellipsoids automatically.

To run the program, we perform:

$./ellipsoid -isTransX <0/1/2> -withEllipsoid <0/1> -initwithTransE <0/1> -input <dir> -output <dir> -size <int> -sizeR <int> -margin_E <float> -margin_R <float> -margin_S <float> -epochs_E <int> -epochs_R <int> -epochs_S <int> -alpha <float> -alpha <float> -rate <float> -rate <float> -note <string>

parameters:

-isTransX <0/1/2>: 1: train TransR and use the results to train ellipsoids. 2: train STransE and use the results to train ellipsoids. 0: only train TransE and use the results to train ellipsoid.

-withEllipsoid <0/1>: 0: not train ellipsoids. 1: train ellipsoids.

-initwithTransE <0/1>: 0: not init with TransE. 1: init with TransE ( when training TransR/STransE)

-input <dir>: path to the dataset directory.

-output <dir>: path to the output directory.

-size <int>: the number of entity vector dimensions.

-sizeR <int>: the number of relation vector dimensions.

-margin_E <float>: the margin hyper-parameter for TransE.

-margin_R <float>: the margin hyper-parameter for TransR.

-margin_S <float>: the margin hyper-parameter for STransE.

-epochs_E <int>: the epochs for TransE.

-epochs_R <int>: the epochs for TransR.

-epochs_S <int>: the epochs for STransE.

-alpha <float>: the SGD learning rate for TransE, TransR and STransE.

-rate <float>: the SGD learning rate for ellipsoid.

-thread <int>: the number of worker threads.

-note <string>: information you want to add in the filename.

test

For evaluating link/entity prediction, the program provides meanrank, Hits@1, Hits@5 and Hits@10 in two setting protocols "Raw" and "Filtered". The testing codes are divided to 3 files, test_ellipsoid.cpp, test_R_ellipsoid.cpp and test_ST_ellipsoid.cpp respectively. Each for testing DRE(TransE), DRE(TransR) and DRE(STransE).

$ ./test_ellipsoid -domain <0/1> -size <int> -sizeR <int> -input <dir> -init <dir> -load <dir> -thread <int> -note <string>

$ ./test_R_ellipsoid -domain <0/1> -size <int> -sizeR <int> -input <dir> -init <dir> -load <dir> -thread <int> -note <string>

$ ./test_ST_ellipsoid -domain <0/1> -size <int> -sizeR <int> -input <dir> -init <dir> -load <dir> -thread <int> -note <string>

parameters:

-domain <0/1>: 0: test without domain; 1: test with domain.

-size <int>: the number of entity vector dimensions.

-sizeR <int>: the number of relation vector dimensions.

-input <dir>: path to the dataset directory.

-init <dir>: path to the baseline model vectors directory.

-load <dir>: path to the ellipsoid vectors directory.

-thread <int>: the number of worker threads.

-note <string>: information you want to add in the filename. (_TransE/_TransR/_STransE are needed for regular output by ellipsoid.cpp)

Test Results

After 10/3/1: , the first four lines are for overall results; 5-8 lines are for 1-to-1 relations; 9-12 lines are for 1-to-N results; 13-16 lines are for N-to-1 results ; 17-20 lines are for N-to-N results; 21-24 lines are for type constraint results

domain-representation-for-knowledge-graph-embedding's Issues

when to release the source code

I am interested in your work, and hope to see the source code, thanks.