GRU4REC-PyTorch

• PyTorch Implementation of the GRU4REC model.
• Original paper: Session-based Recommendations with Recurrent Neural Networks(ICLR 2016)
• This code is based on pyGRU4REC that is implemented by Younghun Song (yhs-968) and original Theano code written by the authors of the GRU4REC paper
• In this version, I added TOP1-max, BPR-max, restructure project, init weight and load model to evaluate.

Requirements

• PyTorch 0.4.1
• Python 3.5
• pandas
• numpy 1.14.5

Usage

Pre processing data

You need to run preprocessing.py to obtain training data and testing data. In this code, I created the mode that extend x percent of last training data to improve result.

The format of data is:

• Filenames
  • Training set should be named as rsc15_train_full.txt
  • Test set should be named as rsc15_test.txt
• Contents
  • rsc15_train_full.txt, rsc15_test.txt should be the tsv files that stores the pandas dataframes that satisfy the following requirements(without headers):
    • The 1st column of the file should be the integer Session IDs
    • The 2nd column of the file should be the integer Item IDs
    • The 3rd column of the file should be the Timestamps

Training and Testing

The project have a structure as below:

├── GRU4REC-pytorch
│   ├── checkpoint
│   ├── data
│   │    ├── preprocessed_data
│   │    │    ├── rsc15_test.txt
│   │    │    ├── rsc15_train_full.txt
│   │    │    ├── rsc15_train_tr.txt
│   │    │    ├── rsc15_train_valid.txt
│   │    ├── raw_data
│   │    │    ├── yoochoose-clicks.dat
│   ├── lib
│   ├── main.py
│   ├── preprocessing.py
│   ├── tool.py

I use tool.py to get 1/8 last yoochoose-clicks.dat

In GRU4REC-pytorch

Training

python3 main.py

Testing

python3 main.py --eval --load_model checkpoint/11081713/model_00004.pt

Logs

PARAMETER----------
BATCH_SIZE=50
CHECKPOINT_DIR=checkpoint/11081713
CUDA=True
DATA_FOLDER=data/preprocessed_data_FULL
DROPOUT_HIDDEN=0.5
DROPOUT_INPUT=0
EMBEDDING_DIM=-1
EPS=1e-06
FINAL_ACT=tanh
HIDDEN_SIZE=100
IS_EVAL=False
LOAD_MODEL=None
LOSS_TYPE=BPR
LR=0.01
MODEL_NAME=GRU4REC
MOMENTUM=0
N_EPOCHS=10
NUM_LAYERS=1
OPTIMIZER_TYPE=Adagrad
SAVE_DIR=models
SEED=7
SIGMA=None
TEST_DATA=rsc15_test.txt
TIME_SORT=False
TRAIN_DATA=rsc15_train_full.txt
VALID_DATA=rsc15_test.txt
WEIGHT_DECAY=0
-------------------
/usr/local/lib/python3.5/dist-packages/torch/nn/modules/rnn.py:38: UserWarning: dropout option adds dropout after all but last recurrent layer, so non-zero dropout expects num_layers greater than 1, but got dropout=0.5 and num_layers=1
  "num_layers={}".format(dropout, num_layers))
Epoch: 0, loss: 0.34, recall: 0.57, mrr: 0.24, time: 2962.893368244171
Save model as checkpoint/11081713/model_00000.pt
Epoch: 1, loss: 0.32, recall: 0.60, mrr: 0.25, time: 2961.867926120758
Save model as checkpoint/11081713/model_00001.pt
Epoch: 2, loss: 0.32, recall: 0.60, mrr: 0.25, time: 2961.7242798805237
Save model as checkpoint/11081713/model_00002.pt
Epoch: 3, loss: 0.31, recall: 0.61, mrr: 0.25, time: 2960.437418460846
Save model as checkpoint/11081713/model_00003.pt
Epoch: 4, loss: 0.31, recall: 0.61, mrr: 0.25, time: 2958.951169013977
Save model as checkpoint/11081713/model_00004.pt
Epoch: 5, loss: 0.31, recall: 0.61, mrr: 0.25, time: 2961.2622771263123
Save model as checkpoint/11081713/model_00005.pt
Epoch: 6, loss: 0.31, recall: 0.61, mrr: 0.25, time: 2961.8305492401123
Save model as checkpoint/11081713/model_00006.pt
Epoch: 7, loss: 0.31, recall: 0.61, mrr: 0.25, time: 2961.3042261600494
Save model as checkpoint/11081713/model_00007.pt
Epoch: 8, loss: 0.30, recall: 0.61, mrr: 0.25, time: 2961.863669157028
Save model as checkpoint/11081713/model_00008.pt
Epoch: 9, loss: 0.30, recall: 0.61, mrr: 0.25, time: 2962.760101079941
Save model as checkpoint/11081713/model_00009.pt

Results

With loss function BPR and full dataset, I get 0.61 in recall@20 and 0.25 in mrr@20

With loss function TOP1 and full dataset, I get 0.62 in recall@20 and 0.26 in mrr@20