OpenHGNN/openhgnn/dataset

• Dataset
• • NodeClassificationDataset
  • • • RDF_NodeCLassification
      • HIN_NodeCLassification
      • • ACM
        • academic4HetGNN
        • DBLP
        • IMDB
      • OGB_NodeClassification
      • • ogbn-mag
  • LinkPredictionDataset
  • • • HIN_LinkPrediction
      • • academic4HetGNN
      • KG_LinkPrediction
  • RecommendationDataset
  • How to build a new dataset
  • • Overview
    • • The Flow

Dataset

A dataset is related to a task, so it is a part of task.

So dataset should load not only a heterograph[DGLGraph], but also some index involving training, validation and testing. In OpenHGNN, we preprocess the feature of dataset outside of model. Specifically, we use a linear layer with bias for each node type to map all node features to a shared feature space. And for no feature nodes, we give a embedding as its feature. Refer to HeteroFeature.

NodeClassificationDataset

• RDF_NodeCLassification

  • AIFB/MUTAG/BGS/AM

• HIN_NodeCLassification

  • ACM

    	author	paper	Subject	Paper-Author	Paper-Subject	Features	Train	Val	Test
    acm4GTN	5,912	3,025	57	9,936	3,025	1,902	600	300	2,125
    acm_han_raw	17,351	4,025	72	13,407	4,025	1,903	808	401	2,816
    acm4NSHE	7,167	4,019	60	13,407	4,019	128(Embedding from deep walk)	-	-	-

  • academic4HetGNN

    	author	paper	Venue	Paper-Author	Paper-venue	Paper-paper
    academic4HetGNN	28,646	21,044	18	69,311	21,044	21,357

  • DBLP

    	author	paper	Conf	Venue	Paper-Author	Paper-Conf	Paper-Term	Train	Val	Test
    dblp4HAN	4,057	14,328	20	8,789	19,645	14,328	88,420	800	400	2,857
    dblp4GTN
    dblp4MAGNN	4,057	14,328	20	7,723	19,645	14,328	85,810	400	400	3257

  • IMDB

    	Movie	Actor	Director	Movie-Actor	Movie-Director	Train	Val	Test
    imdb4HAN	4,780	5,841	2,269	14,340	4,780	300	300	2,687
    imdb4GTN	4,661	5,841	2,270	13,983	4,661	300	300	2,339
    imdb4MAGNN	4,278	5,257	2,081	12,828	4,278	400	400	3,478

• **HGBn **

  The datasets are HGB for Node Classification

  Note：The test data labels are randomly replaced to prevent data leakage issues, refer to HGB.

  In OpenHGNN, you will get the test results in ./openhgnn/output/{model_name}/. If you want to obtain test scores, you need to submit your prediction to HGB's website.

  • HGBn-ACM

  paper	author	subject	term	paper-author	paper-paper	paper-subject	paper-term	Val	Test
  3025	5959	56	1902	9949	5343	3025	255619	907	2118

  • HGBn-IMDB

  movie	actor	director	keyword	actor-movie	director-movie	keyword-movie	train	test
  4932	6124	2393	7971	14779	4932	23610	1371	3202

  • HGBn-Freebase: no feature

  BOOK	BUSINESS	FILM	LOCATION	MUSIC	ORGANIZATION	PEOPLE	SPORTS	train	test
  40402	7153	19427	9368	82351	2731	17641	1025	2386	5568

  • HGBn-DBLP

  author	paper	term	venue	author-paper	paper-term	paper-venue	train	test
  4057	14328	7723	20	19645	85810	14328	1217	2840

• OGB_NodeClassification

  • ogbn-mag

    • Train 629,571 predict venue labels of all papers published before 2018
    • Validation: 64879 papers published in 2018
    • Test: 41939 papers published since 2019

LinkPredictionDataset

• HIN_LinkPrediction

  • academic4HetGNN

• HGBl

  The datasets are HGB for Link Prediction.

  Note：The test data labels are randomly replaced to prevent data leakage issues, refer to HGB.

  In OpenHGNN, you will get the test results in ./openhgnn/output/{model_name}/. If you want to obtain test scores, you need to submit your prediction to HGB's website.

  • HGBl-amazon

    	product	features	product-product0	product-product1	test : product-product0	test : product-product1
    HGBl-amazon	10099	1156	76924	71735	7609	7137

  • HGBl-LastFM

    	user	artist	tag	feature	user-artist	user-user	artist-tag	test:user-artist
    HGBL-LastFM	1892	17632	1088	0	92834	25434	23253	18567

  • HGBl-PubMed

    	node0	node1	node2	node3	feature	node0- node0	node0-node1	node1-node1	node2-node0	node2-node1	node2-node2	node2-node3	node3-node0	node3-node1	node3-node2	test:node1-node1
    HGBL-LastFM	13168	19507	25691	2783	200	16105	25962	42637	31277	51323	62187	6297	3155	5245	798	8528

• KG_LinkPrediction

  • 'wn18', 'FB15k', 'FB15k-237'

RecommendationDataset

• Amazon

  (Containing rating and timestamp information)

  (Source : http://jmcauley.ucsd.edu/data/amazon/)

  Edata['rate'] in user-item edge is the rating.

  It addresses the two most common scenarios in collaborative filtering:

  • rating prediction (e.g. on a scale of 1 to 5 stars), and
  • item prediction from positive-only feedback.

  	User	Item	View	Category	Brand	User-Item	Item-View	Item-Category	Item-Brand	Test(20%) User-Item
  Amazon	6,170	2,753	3,857	22	334	195,791	5,694	5,508	2,753	39,159

• MTWM

  	user	poi	Sup		poi-contain-spu	user-buy-poi	user-buy-spu	user-click-poi
  MTWM	188,155	3,474	16,889		92,024	542,915	1,797,283	1,477,316

How to build a new dataset

Overview

We use dgl.heterograph as our graph data structure.

The API dgl.save_graphs and dgl.load_graphs can be used in storing graph into the local.

The Flow

1. Process your dataset as dgl.heterograph.
2. Store as graph.bin. Compress as dataset_name4model_name.zip
3. Upload the zip file to s3.
4. If the dataset is Heterogeneous Information Network, you can modify the AcademicDataset directly. Or you can refer to it building a new Class Dataset.

We give a demo to build a new dataset.