DGRec-pytorch

This repository implements DGRec proposed in "Session-Based Social Recommendation via Dynamic Graph Attention Network (WSDM 2019)" using PyTorch. We refer to the original code (link) which was implemented in Tensorflow.

This repository is developed by Taesoo Kim, and commented by Jinhong Jung.

Getting Started

Installation

This project requires Python 3.8 and the following Python libraries:

• numpy == 1.22.0
• pandas == 1.3.5
• matplotlib == 3.5.1
• torch == 1.8.2
• fire == 0.4.0
• tqdm == 4.62.3
• loguru == 0.5.3
• dotmap == 1.3.17

To install the pacakges used in this repository, type the following command at the project root:

pip install -r requirements.txt

Data Preparation

Before running the model, you need to prepare datasets to be trained as follows:

unzip datasets/$DATASET.zip -d datasets/$DATASET/

where $DATASET is one of bookdata, musicdata and moviedata.

How To Run

You can simply check if the model works correctly with the following command:

PYTHONPATH=src python3 run.py --data_name $DATASET

The above command will start learning the model on the $DATASET with the specified parameters saved in param.json.

Usage

To use those scripts properly, move your working directory to ./src.

You can tune the hyperparameters and run this project to simply type the following in your terminal:

python3 -m main_trainer \
        --model ··· \
        --data_name ··· \
        --seed ··· \
        --epochs ··· \
        --act ··· \
        --batch_size ··· \
        --learning_rate ··· \
        --embedding_size ··· \
        --max_length ··· \
        --samples_1 ··· \
        --samples_2 ··· \
        --dropout ··· \
        --decay_rate ... \
        --gpu_id ... \

Arguments

Arguments	Explanation	Default
model	Name of model	'DGRec'
data_name	Name of data	'bookdata'
seed	Random seed	0
epochs	Number of traning epochs	20
act	Type of activation function	'relu'
batch_size	Size of batch	100
learning_rate	Learning rate of model	0.002
embedding_size	Size of item and user embedding	50
max_length	Max item count reflected in each user_embedding	20
samples_1	Number of target user's friends	10
samples_2	Number of target user's friends' friends	5
dropout	Dropout rate	0.2
decay_rate	weight decay rate	0.98
gpu_id	Id of gpu you use	0

Repository Structure

The overall file structure of this repository is as follows:

DGRec-pytorch
    ├── README.md
    ├── datasets                        
    ├── requirments
    ├── run.py                          # starts training the model with specified hyperparameters
    └── src         
        ├── utils.py                    # contains utility functions such as setting random seed and showing hyperparameters
        ├── data.py                     # loads a specified dataset
        ├── main.py                     # processes input arguments of a user for training
        └── models                      
            ├── __init__.py
            ├── eval.py                 # evaluates the model with a validation dataset
            ├── model.py                # implements the forward function of the model
            ├── train.py                # implements a function for training the model with hyperparameters
            └── batch
                ├── __init__.py
                ├── minibatch.py        # splits a dataset to mini-batches
                └── neigh_sampler.py    # samples neighborhoods of a given node

Data

Input Data Files

The datasets are from the original repository (link). The format of each file is as follows:

• train.tsv: includes historical behaviors of users. The data in the file is loaded into pandas.Dataframe with five fields such as (SessionId UserId ItemId Timestamps TimeId)
• valid.tsv: the same format as train.tsv, used for tuning hyperparameters.
• test.tsv: the same format as train.tsv, used for testing model.
• adj.tsv: an edge list representing relationships between users, which is also organized by pandas.Dataframe in two fields (FromId, ToId).
• latest_session.tsv: serves as 'reference' to target user. This file records all users available session at each time slot. For example, at time slot t, it stores user u's t-1 th session.
• user_id_map.tsv: maps original string user name to integer id.
• item_id_map.tsv: maps original string item name to integer id.

Data Statistics

The statistics of bookdata, musicdata, and moviedata from the Douban domain are summarized as follows:

Dataset	#user	#item	#event
bookdata	46,548	212,995	1,908,081
musicdata	39,742	164,223	1,792,501
moviedata	94,890	81,906	11,742,260

Experiments

We compare our implmentation compared to the original one in terms of recall@20 and ndcg. We report average metrics with thier standard deviations of 10 runs.

Recommendation accuracies on test sets using the original version in Tensorflow

data	recall@20	ndcg
bookdata	0.3771	0.2841
musicdata	0.3382	0.2539
moviedata	0.1861	0.1950

Recommendation accuracies on test sets using our version in Pytorch

data	recall@20	ndcg
bookdata	0.4034 ± (0.0160)	0.3140 ± (0.0059)
musicdata	0.3777 ± (0.0164)	0.2947 ± (0.0113)
moviedata	0.1594 ± (0.0031)	0.1955 ± (0.0015)

• Note that the above results were obatined with the early-stopping technqiue w.r.t. ndcg, i.e., if you change the base metric of the early-stopping to recall@20, the test recall of our version can be improved as similar as that of the original version.

Hyperparameter tuning

The following tables summarize the experimental results to find better hyperparameters for each dataset.

Results on bookdata (batch_size : 100)

We first fixed the batch size to 100, and varied the size of embeddings as follows:

Embedding size	recall@20	ndcg
100	0.3673 ± (0.0261)	0.3035 ± (0.0130)
50	0.3718 ± (0.0250)	0.3066 ± (0.0116)
30	0.3551 ± (0.0431)	0.2998 ± (0.0132)

The table shows that when the embedding size is 50, the model produces the best accuracy among the results. Then, we fixed the embedding size to 50, and checked the effects of activation functions, drop-out, learning rate, and decay rate as follows:

Activation	Drop-out	Learning rate	Decay rate	recall@20	ndcg
relu	0.2	0.002	0.99	0.3449 ± (0.0181)	0.3002 ± (0.0091)
relu	0.2	0.002	0.99	0.2904 ± (0.0172)	0.2705 ± (0.0092)
relu	0.2	0.002	0.98	0.1894 ± (0.0233)	0.2239 ± (0.0039)
relu	0.2	0.01	0.95	0.3877 ± (0.0341)	0.3106 ± (0.0083)
relu	0.2	0.01	0.99	0.3911 ± (0.0271)	0.3166 ± (0.0091)
relu	0.2	0.01	0.98	0.3705 ± (0.0219)	0.3089 ± (0.0090)
relu	0.3	0.002	0.95	0.3391 ± (0.0193)	0.2999 ± (0.0071)
relu	0.3	0.002	0.98	0.2928 ± (0.0159)	0.2681 ± (0.0055)
relu	0.3	0.002	0.95	0.1857 ± (0.0229)	0.2220 ± (0.0064)
relu	0.3	0.01	0.99	0.3923 ± (0.0211)	0.3062 ± (0.0064)
relu	0.3	0.01	0.98	0.3862 ± (0.0327)	0.3092 ± (0.0084)
relu	0.3	0.01	0.95	0.3724 ± (0.0359)	0.3092 ± (0.0083)
elu	0.2	0.002	0.99	0.3456 ± (0.0274)	0.2967 ± (0.0130)
elu	0.2	0.002	0.99	0.2925 ± (0.0175)	0.2711 ± (0.0107)
elu	0.2	0.002	0.98	0.1946 ± (0.0259)	0.2238 ± (0.0060)
elu	0.2	0.01	0.95	0.3881 ± (0.0224)	0.3101 ± (0.0057)
elu	0.2	0.01	0.99	0.3930 ± (0.0286)	0.3149 ± (0.0073)
elu	0.2	0.01	0.98	0.3561 ± (0.0172)	0.3057 ± (0.0084)
elu	0.3	0.002	0.95	0.3473 ± (0.0199)	0.3014 ± (0.0068)
elu	0.3	0.002	0.98	0.2960 ± (0.0150)	0.2697 ± (0.0085)
elu	0.3	0.002	0.95	0.1818 ± (0.0211)	0.2200 ± (0.0063)
elu	0.3	0.01	0.99	0.4034 ± (0.0160)	0.3140 ± (0.0059)
elu	0.3	0.01	0.98	0.3901 ± (0.0262)	0.3116 ± (0.0050)
elu	0.3	0.01	0.95	0.3705 ± (0.0266)	0.3073 ± (0.0079)

Results on musicdata (batch_size : 50)

We first fixed the batch size to 100, and varied the size of embeddings as follows:

Embedding size	recall@20	ndcg
100	0.3529 ± (0.0295)	0.2962 ± (0.0103)
50	0.3584 ± (0.0181)	0.2876 ± (0.0121)
30	0.3327 ± (0.0259)	0.2709 ± (0.010s9)

The table shows that when the embedding size is 100, the model produces the best accuracy among the results. Then, we fixed the embedding size to 100, and checked the effects of activation functions, drop-out, learning rate, and decay rate as follows:

Activation	Drop-out	Learning rate	Decay rate	recall@20	ndcg
relu	0.2	0.002	0.99	0.3310 ± (0.0229)	0.2772 ± (0.0074)
relu	0.2	0.002	0.99	0.2990 ± (0.0257)	0.2597 ± (0.0082)
relu	0.2	0.002	0.98	0.2184 ± (0.0181)	0.2147 ± (0.0066)
relu	0.2	0.01	0.95	0.3791 ± (0.0246)	0.2969 ± (0.0122)
relu	0.2	0.01	0.99	0.3777 ± (0.0164)	0.2947 ± (0.0113)
relu	0.2	0.01	0.98	0.3523 ± (0.0177)	0.2868 ± (0.0064)
relu	0.3	0.002	0.95	0.3389 ± (0.0239)	0.2808 ± (0.0083)
relu	0.3	0.002	0.98	0.2902 ± (0.0258)	0.2565 ± (0.0141)
relu	0.3	0.002	0.95	0.2215 ± (0.0134)	0.2130 ± (0.0058)
relu	0.3	0.01	0.99	0.3578 ± (0.0183)	0.2956 ± (0.0100)
relu	0.3	0.01	0.98	0.3694 ± (0.0116)	0.2947 ± (0.0096)
relu	0.3	0.01	0.95	0.3516 ± (0.0263)	0.2863 ± (0.0100)
elu	0.2	0.002	0.99	0.3333 ± (0.0371)	0.2712 ± (0.0159)
elu	0.2	0.002	0.99	0.2849 ± (0.0247)	0.2506 ± (0.0131)
elu	0.2	0.002	0.98	0.2110 ± (0.0189)	0.2132 ± (0.0069)
elu	0.2	0.01	0.95	0.3704 ± (0.0181)	0.2914 ± (0.0081)
elu	0.2	0.01	0.99	0.3742 ± (0.0273)	0.2926 ± (0.0073)
elu	0.2	0.01	0.98	0.3601 ± (0.0289)	0.2877 ± (0.0126)
elu	0.3	0.002	0.95	0.3200 ± (0.0292)	0.2725 ± (0.0113)
elu	0.3	0.002	0.98	0.2845 ± (0.0243)	0.2482 ± (0.0117)
elu	0.3	0.002	0.95	0.2116 ± (0.0232)	0.2111 ± (0.0101)
elu	0.3	0.01	0.99	0.3911 ± (0.0265)	0.2886 ± (0.0120)
elu	0.3	0.01	0.98	0.3694 ± (0.0236)	0.2915 ± (0.0096)
elu	0.3	0.01	0.95	0.3544 ± (0.0279)	0.2801 ± (0.0104)

Results on moviedata (batch_size : 500)

Since the dataset is the largest among tested data, we checked the effects of only embedding size and learning rate as follows (see other hyperparameters in here):

Embedding size	Learning rate	recall@20	ndcg
100	0.002	0.1594 ± (0.0031)	0.1955 ± (0.0015)
100	0.01	0.1574 ± (0.0039)	0.1912 ± (0.0024)
50	0.002	0.1509 ± (0.0021)	0.1885 ± (0.0009)
50	0.01	0.1586 ± (0.0033)	0.1901 ± (0.0023)
50	0.05	0.0921 ± (0.0080)	0.1473 ± (0.0040)
30	0.002	0.1349 ± (0.0029)	0.1748 ± (0.0014)

References

[1] Song, W., Xiao, Z., Wang, Y., Charlin, L., Zhang, M., & Tang, J. (2019, January). Session-based social recommendation via dynamic graph attention networks. In Proceedings of the Twelfth ACM international conference on web search and data mining (pp. 555-563).