Reinforced Anchor Knowledge Graph Generation for News Recommendation Reasoning

This repository contains the source code of the paper: Reinforced Anchor Knowledge Graph Generation for News Recommendation Reasoning

Dataset:

The original data we used is from the public news dataset : MIND. We build an item2item dataset based on the method in the paper.

Files in data folder:

• ./data/
  • kg/wikidata-graph/
    • wikidata-graph.tsv knowledge graph triples from Wikidata
    • entity2id.txt entity label to index
    • relation2id.txt relation label to index
    • entity2vecd100.vec entity embedding from TransE
    • relation2vecd100.vec relation embedding from TransE
  • mind/
    • behaviors.tsv the impression logs and users' news click hostories
    • news.tsv the detailed information of news articles involved in the behaviors.tsv file
  • item2item/
    • all_news.tsv all news used for training, validating, testing
    • doc_feature_embedding.tsv document embedding from sentence-bert
    • doc_feature_entity.tsv entities mentioned in documents
    • pos_train.tsv positive item pairs in train data
    • pos_valid.tsv positive item pairs in valid data
    • pos_test.tsv positive item pairs in test data
    • random_neg_sample_train.tsv item2item train data
    • random_neg_sample_valid.tsv item2item valid data
  • kprn/
    • train_data.json train data for KPRN
    • valid_data.json valid data for KPRN
    • predict_train.json warm up train data for anchorKG
    • predict_valid.json warm up valid data for anchorKG

Requirements:

python == 3.9.13
torch == 1.12.0
sklearn == 1.1.2
numpy == 1.23.4
hnswlib == 0.4.0
networkx == 2.8.7
nni == 2.8
sentence_transformers == 2.2.2
tqdm == 4.64.1

How to run the code:

1. Dataset download and process

   $ python data_process.py

   The config file is ./config/data_config.json

   If the download speed is too slow, you can refer to followng links for dataset download and put it under the corresponding folder before running the code.

   • MIND_large_train: ./data/mind/train/
   • MIND_large_valid: ./data/mind/valid/
   • MIND_small_train: ./data/mind/train/
   • MIND_small_valid: ./data/mind/valid/
   • Knowledge Graph: ./data/kg/

2. Kprn training

   $ python KPRN_train.py

   The config file is ./config/KPRN_config.json

3. Warmup training + AnchorKG training

   $ python main.py

   The config file is ./config/anchorkg_config.json

Automatic hyper-parameter tuning

We integrates with NNI module for tuning the hyper-parameters automatically. You can tune the KPRN training stage, warmpup training stage, anchorKG training stage respectively. For easy usage, you can run the following code:

$ nnictl create --config ./nni_config.yaml --port 9074

You can configure the nni_config.yaml for your own usage. For more details about NNI, please refer to NNI Documentation