This repository is that we use the mothod HPE[1] on OGB Link Prediction Task. Especially, we focus on the dataset ogbl-citation2. If you want to learn more about HPE, please click here.

In order to implement the codes in this repository, you have to install the Python package listed below first:

• ogb == 1.3.1
• torch >= 1.6

Or, you can install required Python packages via requirements.txt.

pip3 install -r requirements.txt

Furthermore, the main method HPE is contained in the repository smore, so you have to follow the steps in next section to get and compile the module.

Our method HPE is contained in the repository smore, but when you clone this repository, there would be nothing in the smore repository. Because of this, we have to follow the commands shown below to get the submodue smore.

git submodule init
git submodule update

After the steps above, we can successfully get the whole codes in smore.

Because the module in smore is written in c++, we have to compile the program before using these codes. Therefore, we should go the the root of this repository and compile via the following command:

make

In this compilation process, we can decompose it into two parts:

• Native smore program; because we would like to use the HPE module in this project.
• Content-based initialized HPE program; because we would like to use the pre-trained features provided by dataset to initial node embedding to train the HPE model.

After compilation, there are execution files, HPE, generated. Moreover, there are BPR, HOP-REC, and WARP execution files generated also. You can try these methods which contains in SMORe project if you want.

To use HPE in smore module, first we have to generate the connection network of our dataset. The format of the network should be like this:

userA itemA 3
userA itemC 5
userB itemA 1
userB itemB 5
userC itemA 4

In each row, it means a user select an item, and the number is the weight of this relationship.

In order to generate the network file, we have to run SampleEdges.py. There are three options:

• --dataset: The dataset we want to generate the network. Default dataset is ogbl-citation2.
• --train_network: The file name which we want to save this network. Default is network.txt.
• --train_percent: The number of percentage of edges we want to train. Default is 100.
• --directed: The edges is directed or not. If it is directed, the source node would be add a tag "S" and the target node would be add a tag "T". Default is False.

We set --train_percent as 100 and others as default, so we can run SampleEdges.py as following:

python3 SampleEdges.py --dataset ogbl-citation2 --train_network network.txt --train_percent 100

We use cli constructed in smore to train the network with HPE method. The command is like this:

hpe -train network.txt -embed node-feature.txt -save hpe_rep.txt -dimensions 64 -undirected 0 -sample_times 1200 -walk_steps 5 -threads 8

The command above means that we train the network, which saved in network.txt, using HPE with 1200 sample times and generate the 64-dimensional embedding for each node. Then we save these embeddings to the file hpe_rep.txt. The format in this file should be like this:

6 64
userA 0.0815412 0.0205459 0.288714 0.296497 0.394043 ...
itemA -0.207083 -0.258583 0.233185 0.0959801 0.258183 ...
itemC 0.0185886 0.138003 0.213609 0.276383 0.45732 ...
userB -0.0137994 -0.227462 0.103224 -0.456051 0.389858 ...
itemB -0.317921 -0.163652 0.103891 -0.449869 0.318225 ...
userC -0.156576 -0.3505 0.213454 0.10476 0.259673 ...

You can adjust the size of -thread according to the capability of you device.

After we generate the node embedding file, we can use it on our prediction task. We would run predict.py to predict the results and get the evaluation score. There are four options:

• --dataset: The dataset of our prediction task. Default is ogbl-citation2.
• --embed: The file which save the embedding of each nodes. Default is hpe_rep.txt.
• --val_percent: The number of percentage of validation set you want to test for the evaluation score. Default is 100.
• --test_percent: The number of percentage of testing set you want to test for the evaluation score. Default is 100.

Suppose we set both --val_percent and test_percent as 20, the command we run predict.py should be like this:

python3 predict.py --dataset ogbl-citation2 --val_percent 20 --test_percent 20 --embed hpe_rep.txt

To get the true evaluation score, which means that you want to predict the whole validation set and testing set, be sure to set --val_percent and --test_percent as 100.

To summarize the above, we construct a shell script file run.sh, which contain our experiment steps. You can run this file as following command to get the evaluation score with HPE method.

sh run.sh

Moreover, you can try to change the arguments we set to see whether the score is better or not.

[1] Chih-Ming Chen, Ming-Feng Tsai, Yu-Ching Lin, and Yi-Hsuan Yang. Query-based Music Recommendations via Preference Embedding