D-HAN

The source code of D-HAN

================= UPDATE ===============

The dir caixin_NS_code_4 is just used for running the code on the two datasets well, the code may be not the lasted version reported on the paper.

This is the source code of D-HAN: Dynamic News Recommendation with Hierarchical Attention Network. However, only the code of three tested datasets is uploaded.

Update: 2022.1.9

python 3.6 is used

NOTE: Since several modules are considered in our method, and many ablation studies have been explored, we do not upload the code of each module, if you are interested in any of them, please contact me.

Start from the run_adressa.py file, it will use interactions.py and utils.py file to process news data, use HAN_DNS_time.py file to access the model, then continue run_adressa.py file to train, evaluate and compute metrics value. Attention.py file contains the self-attention, and time embedding; HAN_DNS_time.py file contains element-level, sentences-level, news-level, HAN model and dynamic negative sampling core code, etc.

We remove redundant files, only the training files of public dataset Adressa is kept. To run the training process, do the following steps:

1. Downloading Adressa dataset from [BUAA drive](https://bhpan.buaa.edu.cn:443/link/EA16980123791A175E3B56F92D93438B Valid Until: 2023-04-19 00:59) or Google drive, note that, the dataset is processed from the full dataset of public Adressa dataset, if you need the original processing file, please contact me.
2. Run this command python run_adressa.py, parameters can be set according to the paper or kept default. Note that the number of negative samples used in the training phase is 3, but when dynamic negative sampling method is adopted, 50 news items are first randomly selected and then DNS sample 3 items from the 50 items.

---

Since the main contributions include: (1) We propose to simultaneously capture different granular information, i.e., sentence-, element-, document- and sequence-level information for news recommendation. (2) We propose to recommend news dynamically by a time-aware document-level attention layer, which incorporates the absolute and relative time information. (3) We propose to incorporate negative sampling into the training process to facilitate model optimization.

You can check the code for their corresponding implementation details, it is very easy to understand. Of course, if you have any questions, please create issues in this repo, and I will respond you as soon as possible.