unKR 0.9.1

An Open Source Library for uncertain Knowledge Reasoning

unKR: An Open Source Toolkit for Uncertain Knowledge Graph Representation Learning

English | 中文

unKR is an open source toolkit for Uncertain Knowledge Graph Representation Learning(UKRL). It is based on the PyTorch Lightning framework to decouple the workflow of the UKRL models in order to implement multiple Uncertain Knowledge Graph Embedding(UKGE) methods, which in turn assist knowledge graph complementation, inference and other tasks. The tool provides code implementations and results of various existing UKGE models, and provides users with detailed technical documentation.

🔖 Overview

（图片待修改）

unKR toolkit is an efficient implementation for Uncertain Knowledge Graph Representation Learning(URKL) based on the PyTorch Lightning framework. It provides a refinement module process that can implement a variety of Uncertain Knowledge Graph Embedding(UKGE) models, including UKG data preprocessing(Sampler for negative sampling), model implementation base module, and model training, validation, and testing modules. These modules are widely used in different UKGE models, facilitating users to quickly construct their own models.

There are nine different models available, divided according to whether they are few-shot models or not. unKR has validated the tool on three datasets with seven different evaluation metrics, and the details of the models will be discussed in the following sections.

unKR core development team will provide long-term technical support for the toolkit, and developers are welcome to discuss the work and initiate questions using issue.

Detailed documentation of the unKR technology and results is available at 📋.

📝 Models

unKR implements nine UKGE methods that partition the model based on whether it is a few-shot model or not. The available models are as below.

Category	Model
Normal model	BEURrE, FocusE, GTransE, PASSLEAF, UKGE, UKGsE, UPGAT
Few-shot model	GMUC, GMUC+

Datasets

unKR provides three different sources of UKG datasets including CN15K, NL27K, and PPI5K. The following table respectively shows the source of the datasets and the number of entities, relationships, and quaternions they contain.

Dataset	Source	Entities	Relations	Quaternions
CN15K	ConceptNet	15000	36	241158
NL27K	NELL	27221	404	175412
PPI5K	STRING	4999	7	271666

Organize the three datasets, here are the three data files common to all models.

train.tsv: All triples used for training and corresponding confidence scores in the format(ent1, rel, ent2, score), one quaternion per line.

val.tsv: All triples used for validation and corresponding confidence scores in the format(ent1, rel, ent2, score), one quaternion per line.

test.tsv: All triples used for testing and corresponding confidence scores in the format(ent1, rel, ent2, score), one quaternion per line.

In UKGE, thesoftlogic.tsvfile is also required.

softlogic.tsv: All triples inferred by PSL rule and corresponding inferred confidence scores in the format(ent1, rel, ent2, score), one quaternion per line.

In GMUC, GMUC+, the following five data files are also needed.

train/dev/test_tasks.json: Few-shot dataset with one task per relation in the format{rel:[[ent1, rel, ent2, score], ...]}. The key of the dictionary is the task name and the values are all the quaternions under the task.

path_graph: All data except training, validation and testing tasks, i.e. background knowledge, in the format(ent1, rel, ent2, score). Each line represents a quaternion.

ontology.csv: Ontology knowledge data required for the GMUC+ model, in the format(number, h, rel, t), one ontology knowledge per line. There are four types of rel, which includes is_A, domain, range, and type.

• c1 is_A c2: c1 is a subclass of c2;
• c1 domain c2: the definition domain of c1 is c2;
• c1 range c2: the value domain of c1 is c2;
• c1 type c2: the type of c1 is c2.

Reproduced Results

unKR uses confidence prediction and link prediction tasks for model evaluation in seven different metrics, MSE, MAE(confidence prediction), Hits@k(k=1,3,10), MRR, MR, WMRR, and WMR(link prediction), with raw and filter settings. In addition, unKR adopts a high-confidence filter(set the filter value to 0.7) method for the evaluation.

Here are the reproduced model results on NL27K dataset using unKR as below. See more results in here.

Confidence prediction

Category	Model	MSE	MAE
Normal model	BEUrRE	0.08920	0.22194
	PASSLEAF_ComplEx	0.02434	0.05176
	PASSLEAF_DistMult	0.02309	0.05107
	PASSLEAF_RotatE	0.01949	0.06253
	UKGElogi	0.02861	0.05967
	UKGElogiPSL	0.02868	0.05966
	UKGErect	0.03344	0.07052
	UKGErectPSL	0.03326	0.07015
	UKGsE	0.12202	0.27065
	UPGAT	0.02922	0.10107
Few-shot model	GMUC	0.01300	0.08200
	GMUC+	0.01300	0.08600

Link prediction (filter on high-confidence test data)

Category	Model	Hits@1	Hits@3	Hits@10	MRR	MR	WMRR	WMR
Normal model	BEUrRE	0.156	0.385	0.543	0.299	488.051	0.306	471.784
	FocusE	0.814	0.918	0.957	0.870	384.471	0.871	379.761
	GTransE	0.222	0.366	0.493	0.316	1377.564	0.319	1378.505
	PASSLEAF_ComplEx	0.669	0.786	0.876	0.741	138.808	0.753	138.477
	PASSLEAF_DistMult	0.627	0.754	0.856	0.707	138.781	0.717	137.864
	PASSLEAF_RotatE	0.687	0.816	0.884	0.762	50.776	0.774	50.194
	UKGElogi	0.526	0.670	0.805	0.622	153.632	0.630	152.314
	UKGElogiPSL	0.525	0.673	0.812	0.623	168.029	0.632	167.344
	UKGErect	0.509	0.662	0.807	0.609	126.011	0.614	124.424
	UKGErectPSL	0.500	0.647	0.800	0.599	125.233	0.604	124.189
	UKGsE	0.038	0.073	0.130	0.069	2329.501	0.069	2288.222
	UPGAT	0.618	0.751	0.862	0.701	69.120	0.708	69.364
Few-shot model	GMUC	0.335	0.465	0.592	0.425	58.312	0.426	58.097
	GMUC+	0.338	0.486	0.636	0.438	45.774	0.438	45.682

🛠️ Deployment

Installation

Step1 Create a virtual environment using Anaconda and enter it.

conda create -n unKR python=3.8
conda activate unKR
pip install -r requirements.txt

Step2 Install package.

• Install from source

git clone https://github.com/seucoin/unKR.git
cd unKR
python setup.py install

• Install by pypi

pip install unKR

Step3 Model training.

python main.py

Parameter Adjustment

In the config file, we provide parameter profiles of the reproduced results, and the following parameters can be adjusted for specific use.

parameters:
  confidence_filter:  #whether to perform high-confidence filtering
    values: [0, 0.7]
  emb_dim:
    values: [128, 256, 512...]
  lr:
    values: [1.0e-03, 3.0e-04, 5.0e-06...]
  num_neg:
    values: [1, 10, 20...]
  train_bs:
    values: [64, 128, 256...]

✉️ Citation

If you find unKR is useful for your research, please consider citing the following paper:

@article{
}

😊 unKR Core Team

Southeast University: Jingting Wang, Tianxing Wu, Shilin Chen, Yunchang Liu, Shutong Zhu, Wei Li, Jingyi Xu, Guilin Qi.

🔎 Reference

• NeuralKG: An Open Source Library for Diverse Representation Learning of Knowledge Graphs. SIGIR 2022 Demo. https://arxiv.org/pdf/2202.12571.pdf
• NeuralKG-ind: A Python Library for Inductive Knowledge Graph Representation Learning. SIGIR 2023 Demo. https://arxiv.org/pdf/2304.14678.pdf
• GitHub：https://github.com/zjukg/NeuralKG