# LSAB

**Repository Path**: kyle-liao/lsab

## Basic Information

- **Project Name**: LSAB
- **Description**: LSAB: User Behavioral Pattern Modeling in Sequential Recommendation by Learning Self-Attention Bias 
- **Primary Language**: Python
- **License**: Not specified
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 4
- **Forks**: 0
- **Created**: 2022-09-05
- **Last Updated**: 2023-12-10

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# LSAB

## Setup

Install the required packages into your python environment:

```
pip install -r requirements.txt
```

This code was tested with `python 3.7.13` on `ubuntu` with `cuda 10.1` and various types of GPUs.

## QuickStart

In order to train lsab, run run.py as follows:

```
python run.py --templates train_sas --dataset_code ml-100k --Biasmode learn
```

This will apply all the options specified in `templates/train_sas.yaml`, and train **LSAB** on **MovieLens 100K** dataset as a result.

You can also apply other bias. For example,

```
python run.py --templates train_bert --dataset_code ml-100k --Biasmode lognormal
```

You can also apply other templates in the `templates` folder. For example,

```
python run.py --templates train_bert --dataset_code ml-100k --Biasmode learn
```

will train **BERT4Rec** model instead of **LSAB**.


## Training

Here is a more detailed explanation of how one can train a model.

We will explain in two levels ('Big' Choices and 'Small' Choices).

Remember that you can always study the template files in `./templates` to learn about desirable choices.
### 'Big' Choices

This project is highly modularized so that any (valid) combination of `model`, `Biasmode`,`dataset`, `dataloader`, `negative_sampler` and `trainer` will run.

#### Model

Currently, this repository provides several other baselines.

Choose one of these for `--model_code` option
* meantime
* sas
* tisas
* bert

#### Biasmode

Choose one of these for `--Biasmode` option
* learn
* abs
* normal
* lognormal

#### Privacy
Choose one of these for `--use_epsilon` option
* 1 when use_epsilon==1, use epsilon
* 0
* 
Privacy parameter, you can adjust it.
sensitivity: 1
epsilon: 0.1



#### Dataset

We experimented with four datasets: **MovieLens 100K**, **MovieLens 1M**, **Amazon Beauty** and **Amazon Game**.

Choose one of these for `--dataset_code` option
* ml-100k
* ml-1m
* beauty
* game

The raw data of these datasets will be automatically downloaded to `./Data` the first time they are required.

They will be preprocessed according to the related hyperparameters and will be saved also to `./Data` for later re-use.

Note that downloading/preprocessing is done only once per every setting to save time.

If you want to change the Data folder's path from `./Data` to somewhere else (e.g. shared folder), modify `LOCAL_DATA_FOLDER` variable in `meantime/config.py`.


#### Negative Sampler

There are two types of negative samplers:
* random (sample by random)
* popular (sample according to item's popularity)

Choose one for `--train_negative_sampler`(used for training) and `--test_negative_sampler`(used for evaluation).

### Results

In the result folder, you will find:

```
.
├── config.json
├── models
│   ├── best_model.pth
│   ├── recent_checkpoint.pth
│   └── recent_checkpoint.pth.final
├── status.txt
└── tables
    ├── test_log.csv
    ├── train_log.csv
    └── val_log.csv
```

Below are the descriptions for the contents of each file.

#### Models

* best_model.pth: state_dict of the best model
* recent_checkpoint.pth: state_dict of the model at the latest epoch
* recent_checkpoint.pth.final: state_dict of the model at the end of the training

#### Tables

* train_log.csv: training loss at every epoch
* val_log.csv: evaluation metrics for validation data at every epoch
* test_log.csv: evaluation metrics for test data at best epoch


## References

The baseline codes were translated to PyTorch from the following repositories:
* **MARANK**: [https://github.com/voladorlu/MARank](https://github.com/voladorlu/MARank)
* **SASRec**: [https://github.com/kang205/SASRec](https://github.com/kang205/SASRec)
* **TiSASRec**: [https://github.com/JiachengLi1995/TiSASRec](https://github.com/JiachengLi1995/TiSASRec)
* **BERT4Rec**: [https://github.com/FeiSun/BERT4Rec](https://github.com/FeiSun/BERT4Rec)
* **BERT4Rec**: [https://github.com/SungMinCho/MEANTIME](https://github.com/SungMinCho/MEANTIME)