Jointly Learning Explainable Rules for Recommendation with Knowledge Graph

Last update: Nov 03, 2022

Related tags

Overview

RuleRec

These are our datasets and implementation for the paper:

Weizhi Ma, Min Zhang, Yue Cao, Woojeong Jin, Chenyang Wang, Yiqun Liu, Shaoping Ma, and Xiang Ren. 2019. Jointly Learning Explainable Rules for Recommendation with Knowledge Graph. In TheWebConf'19.

Please cite our paper if you use our datasets or codes. Thanks!

@inproceedings{ma2019jointly,
  title={Jointly Learning Explainable Rules for Recommendation with Knowledge Graph},
  author={Ma, Weizhi and Zhang, Min and Cao, Yue and Jin, Woojeong and Wang, Chenyang and Liu, Yiqun and Ma, Shaoping and Ren, Xiang},
  booktitle={The World Wide Web Conference},
  pages={1210--1221},
  year={2019},
  organization={ACM}
}

If you have any problem about this work, you can contact Weizhi Ma (mawz12 AT hotmail.com).

RuleRec Datasets

The constructed datasets (two scenarios: Amazon cellphone and Amazon electronic) can be found here, which contain several parts:

Recommendation Data:

train.txt, test.txt: user-item interaction data.

Formatting: 
	user id \t item id

item_dic.txt: A python dic, key = item id in Amazon, value = item id here.

Item Attributes:

title.txt, brand.txt, description.txt: item attributes.

Formatting: 
	item id in Amazon \t the title/brand/description of this item

Item Associations:

also_buy.txt, also_view.txt, buy_after_view.txt, buy_together.txt: item associations.

Formatting:
	item id in Amazon \t items that have also\_buy/also\_view/buy\_after\_view/buy\_together association with this item, split by ' '

Entity Linking Data:

title_entities.txt, brand_entities.txt, description_entities.txt: entity linking results on freebase.

Formatting:
	item id in Amazon \t entity name \t entity id in Freebase

Path data:

KGData/*/rule_score.txt: As Freebase is an extremely large knowledge graph, only the related paths in the knowledge graph are recorded in this file. The head and tail entity of each path linked by at least one item.

training_pairs.txt and usercandidates.txt are two files sampled for rule learning and recommendation. You can replace them with other sampling results. The formatting of training_pairs.txt is 'user id : [positive item id, negative item id]'.

Besides, the original Amazon datasets (including user-item interaction history and item associations) are provided by Professor Mcauley. You can download them here.

Rule Learning Codes

If you want to use these codes, you should download RuleRec dataset and put them together first.

getItemItemDic.py: Enumerate all possible rules.

selectRules.py: Rule selection (rule features for jointly learning will also be generated in this step).

getFeatures.py: Calculate features based on the selected rules for item recommendation.

Environments: Python 3.6.3

sklearn = 0.19.1

numpy = 1.13.3

# Example:
> python getItemItemDic.py Cellphone abu
> python selectRules.py Cellphone abu 50
> python getFeatures.py Cellphone abu 50

RuleRec(BPRMF) Codes:

This implementation is based on MyMediaLiteJava. Both codes and jar file are provided.

The evaluation datasets can be downloaded from here, which is generated from RuleRec Data and contains both rule selection features and rule features.

Environments: Java, version 1.6 or later

# Example 1: Use Cellphone dataset
> java -jar BPRMF.jar --recommender=BPRMF --training-file=./RuleRecInput/Cellphone/trainingSet.txt --test-file=./RuleRecInput/Cellphone/testSet.txt --candidateFile=./RuleRecInput/Cellphone/candidates.txt --trainingPairFile=./RuleRecInput/Cellphone/trainingPairs.txt --trainingFeatures=./RuleRecInput/Cellphone/trainingFeatures.txt --testFeatures=./RuleRecInput/Cellphone/testFeatures.txt --learningRate=0.1 --usermodel=0 --iter-times=30 --rule-weight=0.005  --ruleWeightNumber=200 --resultFile=result.txt 
# output:[email protected]=0.34968 [email protected]=0.48024 [email protected]=0.28287 [email protected] num_users=27840 num_items=100 num_lists=27840

# Example 2: Use Cellphone dataset with jointly learning
> java -jar BPRMF.jar --recommender=BPRMF --training-file=./RuleRecInput/Cellphone/trainingSet.txt --test-./RuleRecInput/Cellphone/testSet.txt --candidateFile=./RuleRecInput/Cellphone/candidates.txt --trainingPairFile=./RuleRecInput/Cellphone/trainingPairs.txt --trainingFeatures=./RuleRecInput/Cellphone/trainingFeatures.txt --testFeatures=./RuleRecInput/Cellphone/testFeatures.txt --learningRate=0.1 --usermodel=0 --iter-times=30 --rule-weight=0.005  --ruleWeightNumber=200 --resultFile=result.txt --trainTogether=2  --lossType=sigmoid --lossCombineRate=0.2 --ruleselectTrain=./RuleRecInput/Cellphone/ruleselect/ --ruleselectResult=./RuleRecInput/Cellphone/ruleselect/ 
# output:[email protected]=0.36430 [email protected]=0.49429 [email protected]=0.29536 [email protected]=0.23214 num_users=27840 num_items=100 num_lists=27840

# Example 3: Use Electronic dataset
> java -jar BPRMF.jar --recommender=BPRMF --training-file=./RuleRecInput/Electronic/trainingSet.txt --test-file=./RuleRecInput/Electronic/testSet.txt --candidateFile=./RuleRecInput/Electronic/candidates.txt --trainingPairFile=./RuleRecInput/Electronic/trainingPairs.txt --trainingFeatures=./RuleRecInput/Electronic/trainingFeatures.txt --testFeatures=./RuleRecInput/Electronic/testFeatures.txt --learningRate=0.05 --ruleWeightNumber=200 --usermodel=0 --iter-times=30 --rule-weight=0.01 --resultFile=result.txt 
# output:[email protected]=0.20694 [email protected]=0.29726 [email protected]=0.17284 [email protected]=0.13483 num_users=18223 num_items=100 num_lists=18223

# Example 4: Use Electronic dataset with jointly learning
> java -jar BPRMF.jar --recommender=BPRMF --training-file=./RuleRecInput/Electronic/trainingSet.txt --test-file=./RuleRecInput/Electronic/testSet.txt --candidateFile=./RuleRecInput/Electronic/candidates.txt --trainingPairFile=./RuleRecInput/Electronic/trainingPairs.txt --trainingFeatures=./RuleRecInput/Electronic/trainingFeatures.txt --testFeatures=./RuleRecInput/Electronic/testFeatures.txt --learningRate=0.05 --ruleWeightNumber=200 --usermodel=0 --iter-times=30 --rule-weight=0.01 --resultFile=result.txt --trainTogether=2  --lossType=sigmoid --lossCombineRate=0.005 --ruleselectTrain=./RuleRecInput/Electronic/ruleselect/ --ruleselectResult=./RuleRecInput/Electronic/ruleselect/ 
# output:[email protected]=0.20798 [email protected]=0.29979 [email protected]=0.17407 [email protected]=0.13570 num_users=18223 num_items=100 num_lists=18223

Jointly Learning Explainable Rules for Recommendation with Knowledge Graph

Related tags

Overview

RuleRec

RuleRec Datasets

Rule Learning Codes

RuleRec(BPRMF) Codes:

Owner

Fast Python Collaborative Filtering for Implicit Feedback Datasets

Plex-recommender - Get movie recommendations based on your current PleX library

Learning Fair Representations for Recommendation: A Graph-based Perspective, WWW2021

Code for ICML2019 Paper "Compositional Invariance Constraints for Graph Embeddings"

Temporal Meta-path Guided Explainable Recommendation (WSDM2021)

Bert4rec for news Recommendation

Persine is an automated tool to study and reverse-engineer algorithmic recommendation systems.

Group-Buying Recommendation for Social E-Commerce

Code for MB-GMN, SIGIR 2021

Approximate Nearest Neighbors in C++/Python optimized for memory usage and loading/saving to disk

A framework for large scale recommendation algorithms.

Real time recommendation playground

Spotify API Recommnder System

6002project-rl - An implemention of offline RL on recommender system

NVIDIA Merlin is an open source library designed to accelerate recommender systems on NVIDIA’s GPUs.

基于个性化推荐的音乐播放系统

Collaborative variational bandwidth auto-encoder (VBAE) for recommender systems.

RetaGNN: Relational Temporal Attentive Graph Neural Networks for Holistic Sequential Recommendation

Self-supervised Graph Learning for Recommendation

Dual Graph Attention Networks for Deep Latent Representation of Multifaceted Social Effects in Recommender Systems