This is our TensorFlow implementation for the paper:
Wang-Cheng Kang, Julian McAuley (2018). Self-Attentive Sequential Recommendation. In Proceedings of IEEE International Conference on Data Mining (ICDM'18)
Please cite our paper if you use the code or datasets.
The code is tested under a Linux desktop (w/ GTX 1080 Ti GPU) with TensorFlow 1.12 and Python 2.
The preprocessed datasets are included in the repo (e.g. data/Video.txt), where each line contains an user id and
item id (starting from 1) meaning an interaction (sorted by timestamp).
The data pre-processing script is also included. For example, you could download Amazon review data from here., and run the script to produce the txt format data.
We crawled reviews and game information from Steam. The dataset contains 7,793,069 reviews, 2,567,538 users, and 32,135 games. In addition to the review text, the data also includes the users' play hours in each review.
- Download: reviews (1.3G), game info (2.7M)
- Example (game info):
{
"app_name": "Portal 2",
"developer": "Valve",
"early_access": false,
"genres": ["Action", "Adventure"],
"id": "620",
"metascore": 95,
"price": 19.99,
"publisher": "Valve",
"release_date": "2011-04-18",
"reviews_url": "http://steamcommunity.com/app/620/reviews/?browsefilter=mostrecent&p=1",
"sentiment": "Overwhelmingly Positive",
"specs": ["Single-player", "Co-op", "Steam Achievements", "Full controller support", "Steam Trading Cards", "Captions available", "Steam Workshop", "Steam Cloud", "Stats", "Includes level editor", "Commentary available"],
"tags": ["Puzzle", "Co-op", "First-Person", "Sci-fi", "Comedy", "Singleplayer", "Adventure", "Online Co-Op", "Funny", "Science", "Female Protagonist", "Action", "Story Rich", "Multiplayer", "Atmospheric", "Local Co-Op", "FPS", "Strategy", "Space", "Platformer"],
"title": "Portal 2",
"url": "http://store.steampowered.com/app/620/Portal_2/"
}To train our model on Video (with default hyper-parameters):
python main.py --dataset=Video --train_dir=default
or on ml-1m:
python main.py --dataset=ml-1m --train_dir=default --maxlen=200 --dropout_rate=0.2
The implemention of self attention is modified based on this
The convergence curve on ml-1m, compared with CNN/RNN based approaches:
