CFKGC_README

Continual Few-shot Knowledge Graph Completion

CIKM 2024: "Learning from Novel Knowledge: Continual Few-shot Knowledge Graph Completion"

This repository contains the implementation of the paper:

Zhuofeng Li, Haoxiang Zhang, Qiannan Zhang, Ziyi Kou, and Shichao Pei. 2024. Learning from Novel Knowledge: Continual Few-shot Knowledge Graph Completion. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management (CIKM '24). Association for Computing Machinery, New York, NY, USA, 1326–1335. https://doi.org/10.1145/3627673.3679734

• Please star our repo 🌟 and cite our paper if you feel useful. Feel free to open an issue if you have any questions.

Introduction

Knowledge graph (KG) completion is vital for uncovering missing knowledge and addressing incompleteness in KGs. While few-shot models following meta-learning have shown promise for rare relations, they typically assume static KGs. In reality, KGs evolve with newly emerging relations, requiring continual learning capabilities.

This work proposes a novel framework for continual few-shot KG completion that addresses two key challenges:

1. Catastrophic forgetting
2. Scarcity of novel relations

Key Features

• Data-level triple rehearsal and model-level meta-learner modulation to combat forgetting
• Multi-view relation augmentation using self-supervised learning
• Enhanced generalization ability for limited novel relations

Method Overview

Getting Started

Installation

git clone https://github.com/cfkgc-paper/CFKGC-paper.git
cd CFKGC-paper
pip install -r requirements.txt

Data Preparation

The experiments use two datasets, the followings provide the embedding we processed:

1. NELL-One [Baidu Netdisk][Google Drive]
2. Wiki-One [Baidu Netdisk][Google Drive]

Total Zip with all codes: [Baidu Netdisk]

The Structure of the project is as followings:

CFKGC-paper
    |--./NELL
    |--./Wiki
    |--trainer.py
    |--params.py
    |--main.py
    |--embedding.py
    |--data_loader.py
    |--...

Statistics of the datasets:

Dataset	#Relation	#Entity	#Triples	#Task
NELL-One	358	68,545	181,109	67
Wiki-One	822	4,838,244	5,859,240	183

Running the Code

The total train and evaluation process with reported hyper-parameters in the paper:

# NELL-One
python main.py -data NELL-One -path ./NELL -if True -br 30 -bs 3 -nt 8 -l 0.1 -es_np 50
# Wiki-One
python main.py -data Wiki-One -path ./Wiki -if False -br 72 -bs 7 -nt 8 -l 1. -es_np 300

Configuration

Extended setting can be set by altering the few-shot size, number-size here are key parameters in params.py:

# Dataset setting
args.add_argument("-data", "--dataset", default="NELL-One", type=str)
args.add_argument("-path", "--data_path", default="./NELL", type=str)

# dataloader setting
# Support Numeber in Base Meta-learn stage
args.add_argument("-bfew", "--base_classes_few", default=3, type=int)
# Query Numeber in Base Meta-learn stage
args.add_argument("-bnq", "--base_classes_num_query", default=3, type=int)
# Support Numeber in Novel Tasks stage
args.add_argument("-few", "--few", default=3, type=int)
# Query Numeber in Novel Tasks stage
args.add_argument("-nq", "--num_query", default=3, type=int)
# Relation Numeber in Base Meta-learn stage
args.add_argument("-br", "--base_classes_relation", default=30, type=int)
# Novel Relations Numeber in each Novel Tasks
args.add_argument("-bs", "--batch_size", default=3, type=int)
# Number of Novel Tasks
args.add_argument("-nt", "--num_tasks", default=8, type=int)

Citation

Please star our repo 🌟 and cite our paper if you feel useful. Feel free to open an issue if you have any questions.

@inproceedings{10.1145/3627673.3679734,
author = {Li, Zhuofeng and Zhang, Haoxiang and Zhang, Qiannan and Kou, Ziyi and Pei, Shichao},
title = {Learning from Novel Knowledge: Continual Few-shot Knowledge Graph Completion},
year = {2024},
isbn = {9798400704369},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
url = {https://doi.org/10.1145/3627673.3679734},
doi = {10.1145/3627673.3679734},
pages = {1326–1335},
numpages = {10},
keywords = {continual learning, few-shot learning, knowledge graphs},
location = {Boise, ID, USA},
series = {CIKM '24}
}