• PyTorch Implementation of the GRU4REC model.
• Original paper: Session-based Recommendations with Recurrent Neural Networks(ICLR 2016)
• Extension over the Original paper: Recurrent Neural Networks with Top-k Gains for Session-based Recommendations(CIKM 2018)
• 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
• This Version supports TOP1, BPR, TOP1-max, BPR-max, and Cross-Entropy Losses.

• PyTorch 0.4.1
• Python 3.5
• pandas
• numpy 1.14.5

RecSys Challenge 2015 Dataset can be retreived from HERE

• You need to run preprocessing.py to obtain training data and testing data. In the paper, only the training set was used, the testing set is ignored.
• The training set itself is divided into training and testing where the testing split is the last day sessions.

The format of data is similar to that obtained from RecSys Challenge 2015:

• Filenames
  • Training set should be named as recSys15TrainOnly.txt
  • Test set should be named as recSys15Valid.txt
• Contents
  • recSys15TrainOnly.txt, recSys15Valid.txt should be the tsv files that stores the pandas dataframes that satisfy the following requirements:
    • The 1st column of the file should be the integer Session IDs with header name SessionID
    • The 2nd column of the file should be the integer Item IDs with header name ItemID
    • The 3rd column of the file should be the Timestamps with header name Time

The project have a structure as below:

├── GRU4REC-pytorch
│   ├── checkpoint
│   ├── data
│   │    ├── preprocessed_data
│   │    │    ├── recSys15TrainOnly.txt
│   │    │    ├── recSys15Valid.txt
│   │    ├── raw_data
│   │    │    ├── yoochoose-clicks.dat
│   ├── lib
│   ├── main.py
│   ├── preprocessing.py
│   ├── tool.py

tool.py can be used to get 1/8 last session from yoochoose-clicks.dat

In GRU4REC-pytorch

Training

python main.py

Testing

python main.py --is_eval --load_model checkpoint/CHECKPOINT#/model_EPOCH#.pt

--hidden_size Number of Neurons per Layer (Default = 100)
--num_layers Number of Hidden Layers (Default = 1)
--batch_size Batch Size (Default = 50)
--dropout_input Dropout ratio at input (Default = 0)
--dropout_hidden Dropout at each hidden layer except the last one (Default = 0.5)
--n_epochs Number of epochs (Default = 10)
--k_eval Value of K used durig Recall@K and MRR@K Evaluation (Default = 20)
--optimizer_type Optimizer (Default = Adagrad)
--final_act Activation Function (Default = Tanh)
--lr Learning rate (Default = 0.01)
--weight_decay Weight decay (Default = 0)
--momentum Momentum Value (Default = 0)
--eps Epsilon Value of Optimizer (Default = 1e-6)
--loss_type Type of loss function TOP1 / BPR / TOP1-max / BPR-max / Cross-Entropy (Default: TOP1-max)
--time_sort In case items are not sorted by time stamp (Default = 0)
--model_name String of model name.
--save_dir String of folder to save the checkpoints and logs inside it (Default = /checkpoint).
--data_folder String of the directory to the folder containing the dataset.
--train_data Name of the training dataset file (Default = recSys15TrainOnly.txt)
--valid_data Name of the validation dataset file (Default = recSys15Valid.txt)
--is_eval Should be used in case of evaluation only using a checkpoint model.
--load_model String containing the checkpoint model to be used in evaluation.
--checkpoint_dir String containing directory of the checkpoints folder.

Different loss functions and different parameters have been tried out and the results can be seen from HERE