LinRec

Source code of LinRec: Linear Attention Mechanism for Long-term Sequential Recommender Systems

All files are conducted in the RecBole framework, and the link to RecBole is https://github.com/RUCAIBox/RecBole.

Environments

The requirements are the same as RecBole as below:
torch>=1.10.0
numpy>=1.17.2
scipy>=1.6.0
hyperopt==0.2.5
pandas>=1.0.5
tqdm>=4.48.2
scikit_learn>=0.23.2
pyyaml>=5.1.0
colorlog==4.7.2
colorama==0.4.4
tensorboard>=2.5.0
thop>=0.1.1.post2207130030
ray>=1.13.0
tabulate>=0.8.10
plotly>=4.0.0
RecBole requires Python version 3.7 or later. RecBole requires torch version 1.7.0 or later. If you want to use RecBole with GPU, please ensure that CUDA or cudatoolkit version is 9.2 or later. This requires NVIDIA driver version >= 396.26 (for Linux) or >= 397.44 (for Windows 10).

Implementations

Then, you may substitute the 'layers.py' file ('recbole/model') in Recbole.

• Config: Other special configuration files (e.g., 'overall.yaml', 'ml-1m.yaml') are included. You must replace them in RecBole ('overall.yaml' -> 'recbole/properties/'; 'ml-1m.yaml' -> 'recbole/properties/dataset/'). Otherwise, errors may occur.

• Data: You have to download the datasets to be used: e.g., ml-100k, ml-1m, gowalla, where we provide the ml-1m for convenience. Then put the downloaded dataset into the 'dataset/' directory in Recbole, like ml-100k.

• Hyperpara: The most critical parameters are $N$ ('MAX_ITEM_LIST_LENGTH') and $d$ ('hidden_size'), which are set in 'ml-1m.yaml' and 'run.py'. Please make sure the values conform to settings in the paper, e.g., $N=200$, $d=128$ for the ml-1m dataset.

Now, You can run the run.py file for a quick start. You can follow the instructions of RecBole for other operations.