• I have heavily modified the code base to work with Python 3.7, as version 3.6.5 seemed to have problems with annotations.
• Fixes to transE originated from hwwang55#29 have been implemented. Credits go to @sinmentis.
• Tensorflow has been updated to version 2.10.0.
• All credit goes to the original authors of the DKN paper and the original repository. This is just a fork to make the code work with newer versions of Python and Tensorflow.

This repository is the implementation of DKN (arXiv):

DKN: Deep Knowledge-Aware Network for News Recommendation
Hongwei Wang, Fuzheng Zhang, Xing Xie, Minyi Guo
The Web Conference 2018 (WWW 2018)

DKN is a deep knowledge-aware network that takes advantage of knowledge graph representation in news recommendation. The main components in DKN is a KCNN module and an attention module:

• The KCNN module is to learn from semantic-level and knowledge-level representations of news jointly. The multiple channels and alignment of words and entities enable KCNN to combine information from heterogeneous sources.
• The attention module is to model the different impacts of a user’s diverse historical interests on current candidate news.

• data/
  • kg/
    • Fast-TransX: an efficient implementation of TransE and its extended models for Knowledge Graph Embedding (from https://github.com/thunlp/Fast-TransX);
    • kg.txt: knowledge graph file;
    • kg_preprocess.py: pre-process the knowledge graph and output knowledge embedding files for DKN;
    • prepare_data_for_transx.py: generate the required input files for Fast-TransX;
  • news/
    • news_preprocess.py: pre-process the news dataset;
    • raw_test.txt: raw test data file;
    • raw_train.txt: raw train data file;
• src/: implementations of DKN.

Note: Due to the pricacy policies of Bing News and file size limits on Github, the released raw dataset and the knowledge graph in this repository is only a small sample of the original ones reported in the paper.

• raw_train.txt and raw_test.txt:
  user_id[TAB]news_title[TAB]label[TAB]entity_info
  for each line, where news_title is a list of words w1 w2 ... wn, and entity_info is a list of pairs of entity id and entity name: entity_id_1:entity_name;entity_id_2:entity_name...
• kg.txt:
  head[TAB]relation[TAB]tail
  for each line, where head and tail are entity ids and relation is the relation id.

The code has been tested running under Python 3.7, with the following packages installed (along with their dependencies):

• tensorflow==2.10.0
• numpy==1.20
• pandas==1.1.5
• scikit-learn==0.24.2
• gensim==4.1.2
• urllib3==1.26.6
• requests==2.27.1
• protobuf==3.19.6
• typing-extensions==4.1.1
• h5py==3.1.0
• wrapt==1.14.1
• tensorboard==2.10.1
• keras==2.10.0
• tensorflow-estimator==2.10.0
• matplotlib==3.5.3

$ pip install -r requirements.txt
$ cd data/news
$ python news_preprocess.py
$ cd ../kg
$ python prepare_data_for_transx.py
$ cd Fast-TransX/transE/ (note: you can also choose other KGE methods)
$ g++ transE.cpp -o transE -pthread -O3 -march=native
$ ./transE
$ cd ../..
$ python kg_preprocess.py
$ cd ../../src
$ python main.py (note: use -h to check optional arguments)