TE4SRec

This repository is the public version of the research project about Temporal Encoding for Sequential Recommendations ( TE4SRec).

All the model networks and training pipelines are based on PyTorch framework.

Quick Start

Start training TE4SRec-b with the following command on Steam dataset.

python main.py --dataset_name steam --is_bert True --max_seq_len 50 --num_worker 1 --train_batch_size 512 --pred_batch_size 128

Usage

To run main.py, there are parameters to be considered. The types, descriptions and default values of these parameters are listed as follows. To change the parameters use --{param_name} {param_value} after main.py.

Parameter	Default Value	Type	Description
seed	42	int	Random seed for reproducibility.
output_path	'./output/results'	str	Path to save the output files.
data_dir	'./data'	str	Directory path for dataset storage.
dataset_name	'yelp'	str	Name of the dataset to be used.
pred_threshold	3	int	Threshold of sequence length for prediction.
is_bert	True	bool	Flag to determine whether BERT or left-to-right attention is used.
model_name	'TE4SRec'	str	Name of the model to be loaded.
load_model	False	bool	Flag to determine if a pre-trained model should be loaded.
max_seq_len	50	int	Maximum sequence length for the model.
alpha	0.5	float	Temporal scaler.
d_temporal	50	int	Dimension of temporal features.
d_model	50	int	Dimension of the model.
num_block	2	int	Number of Transformer blocks in the model architecture.
num_heads	1	int	Number of attention heads in the model.
is_gelu	True	bool	Flag to determine whether GELU or ReLU to be used as activation function.
dropout_rate	0.2	float	Dropout rate for regularization.
mask_ratio	0.2	float	Ratio of masking for BERT training.
pretrain_epoch	0	int	Number of epochs for pre-training without evaluation metrics.
stop_patience	10	int	Patience for early stopping during training.
train_batch_size	512	int	Batch size for training.
num_worker	1	int	Number of workers for data loading in parallel (Torch standard).
lr	0.001	float	Learning rate for the optimizer.
adam_beta1	0.9	float	Beta1 parameter for Adam optimizer.
adam_beta2	0.98	float	Beta2 parameter for Adam optimizer.
is_sampling	False	bool	Flag to enable sampling in evaluation (according to SASRec).
neg_num	100	int	Number of negative samples for evaluation (only for is_sampling=True).
pred_batch_size	128	int	Batch size for prediction.

Dataset Collection

In the project's directory structure, the raw data files are systematically organized under ./data-raw/{dataset_name}. Following preprocessing, the refined data is stored in ./data/{dataset_name}/data.csv. All datasets have been sourced from publicly available resources.

• Steam
• Yelp
• Goodreads
• RateBeer

Data Preprocessing

We preprocess datasets with two core parameters about sampling in this project. The users and items can be sampled randomly to reduce the numbers.

• sample types: ['sample_item', 'sample_user']
• sample ratios

python preprocess.py

Once preprocessing is complete, execute the command below to visualize metadata along with distributions of user-based, item-based, and temporal information. The results are stored in the directory ./output/data/{dataset_name} as default, so make sure the directory existing.

python data_view.py

Project Structure

│   .gitignore
│   dataset.py
│   data_view.py
│   LICENSE
│   main.py
│   preprocess.py
│   README.md
│   result_view.py
│   trainer.py
│   utils.py
│
├───model
│   │   attention.py
│   │   feedforward.py
│   │   TE4SRec.py
│   │   temporal_encoding.py
│
├───data
│   ├───steam
│   │       data.csv
│   └───yelp
│
├───data-raw
│   ├───steam
│   │       steam_new.json
│   └───yelp
│
├───output
│   ├───data
│   │   ├───steam
│   │   │       basic.txt
│   │   │       item_day_gap.png
│   │   │       item_freq_count.png
│   │   │       month_action.png
│   │   │       seq_day_gap.png
│   │   │       seq_len_count.png
│   │
│   └───results
│       ├───steam
│       │   ├───TE4SRec-b
│       │   │       config
│       │   │       eval_valid_line.png
│       │   │       model_checkpoint.pt
│       │   │       temporal_encoding.png
│       │   │       test_metrics.csv
│       │   │       top_k_tables.csv
│       │   │       valid_metrics.csv