brucenccu / ogb_collab_project

OGB_collab_project

Introduction

• This repository mainly applies HPE ,BPR ,WARP, HOP-REC to ogbl-collab dataset whose task is to predict the future author collaboration relationships given the past collaborations.
• More information about models can refer to these papers :

1. HPE : Query-based Music Recommendations via Preference Embedding
2. BPR : BPR: Bayesian personalized ranking from implicit feedback
3. WARP : WSABIE: Scaling Up To Large Vocabulary Image Annotation , Learning to Rank Recommendations with the k-Order Statistic Loss
4. HOP-REC : HOP-Rec: High-Order Proximity for Implicit Recommendation

Requirements

• Python3
• SMORe
• OGB

Quick Start

1. Clone the repository

git clone --recursive https://github.com/brucenccu/OGB_collab_project

or you can execute following code

git clone https://github.com/brucenccu/OGB_collab_project
cd OGB_collab_project
git submodule init
git submodule update

2. Download the dataset and compile the SMORe

cd OGB_collab_project
./prepare.sh

3. If there isn't any error, you can use run.sh to get the prediction. (field_make.py is just for HOPREC)

python3 field_make.py --path ./data/train/train_valid_2010.txt 
./run.sh -model HOPREC -train ./data/train/train_valid_2010.txt -save ./data/train/embed -field ./data/train/field.txt -dim 128 -sample_times 500 -threads 8 -undirected 1

Note : You can get more usage below.

Usage

1. Input data format:

• training data

nodeA nodeB 1
nodeB nodeC 3
nodeA nodeD 2
nodeB nodeE 1
nodeD nodeE 2

• field data (HOP-REC) : to assign the field of each vertex (user or item)

nodeA u
nodeB u
nodeC u
nodeD u 
nodeE u

2. Get the embedding file

a. Initial weights are random

./smore/cli/<model name> -train <input_file> -save <embed_file>

Usage: 
./smore/cli/hpe -train ./train.txt -save ./embed

b. Initial weights are given embedding of nodes.

./smore/cli/<model_name>_node_embed -train <input_file> -save <embed_file> -embed <embed_file> -dimensions <dim>

Usage: 
./smore/cli/hpe_node_embed -train ./train.txt -save ./embed

Parameters:

Options Description:
    -train <string>
        the path of input file
    -save <string>
        Save the representation data
    -embed <string> (only for models whose initial weights are given embeddings of nodes)
        the path of embedding file
    -dimensions <int>
        the dimensions of the input embedding file

Pretrained embedding data format:

//the embedding of the node 0 (0 is the index of the node)
0.1 0.2 -0.4 0.1 0.4 ...
//the embedding of the node 1 
0.2 -0.3 -0.1 0.8 0.4 ...
//the embedding of the node 2 
0.9 0.2 -0.5 0.3 0.1 ...
...

You can get more usages from SMORe.

3. Make prediction to obtain the Hit@50 result.

Run:

python3 predict.py --embed <embed_file> --input_dim <dimensions> --undirected <directions>
python3 logger.py

Parameters:

Options Description:
    --embed <string>
        the path of input embedding file
    --dimension <int>
        the dimensions of the input embedding file
    --undirected <int>
        whether the input graph is undirected or not (1/0)

Example script

Here, you can use the run.sh to run through the whole predict task.

Run:

./run.sh -model <model_name> -train <input_file_path> -save <save_file_path> 

Parameters:

Options Description:
    -model 
        name of the model you want to use.(HPE/BPR/WARP/HOPREC/<model_name>_node_embed)
    -field (optional)
        the path of the field data (HOP-REC)
    -train
        the path of input file
    -save 
        the path od output embedding file
    -embed (optional)
        the path od pretrained node embedding file
    -dim 
        the dimensions of the vertex representation; default is 64
    -sample_times 
        number of training samples *Million; default is 10
    -threads
        number of training threads; default is 1
    -undirected
        whether the input graph is undirected or not; default is 1

Experiment

In addition to using the original dataset, we also filter out edges with a closer year to be the new dataset.The following descriptions are about different datasets.

• train_all : all edges until 2017
• train_2015 : all edges from 2015 - 2017
• train_valid_all : all edges until 2018
• train_valid_2010 : all edges from 2010 - 2018
• train_valid_2015 : all edges from 2015 - 2018

HOPREC for five different datasets

(sample_times = 500, dimensions = 128, alpha = 0.025)

	Without Pretrained (valid/test)	with Pretrained (valid/test)
train_all	0.6588 / 0.5596	0.6639 / 0.5682
train_2015	0.6410 / 0.5426	0.6516 / 0.5551
train_valid_all	1.000 / 0.6687	0.9995 / 0.6805
train_valid_2015	1.000 / 0.6670	1.000 / 0.6741
train_valid_2010	1.000 / 0.7005	1.000 / 0.7077

The result of other models can be referred to this report.