In the PAPER, we pair the ground-truth item with 99 randomly sampled negative items that the user has not interacted with, and report the results of HR@{1, 5, 10}, NDCG@{5, 10} and MRR.
In the repo, we rank the ground-truth item with ALL the items. We omit the FM and AutoInt because they need enumerate all user-item pairs, which take a very long time. The results are shown in the following pic(all_rank.PNG).
Performance comparison of different methods on six datasets. The best performance and the second best performance methods are denoted in bold and underlined fonts respectively.
pip install -r requirements.txt./data/data_process.py
data-name.txt
one user per line
user_1 item_1 item_2 ...
user_2 item_1 item_2 ...
data-name_item2attributes.json
{item_1:[attr, ...], item_2:[attr, ...], ... }python run_pretrain.py \
--data_name Beautypython run_finetune.py \
--data_name Beauty \
--ckp 100