yeganegi-reza/Torch-GRU4Rec

PyTorch implementation GRU4Rec model

Torch-GRU4REC

• PyTorch implementation of the algorithm of Session-based Recommendations with Recurrent Neural Networks(ICLR 2016). With the extensions introduced in Recurrent Neural Networks with Top-k Gains for Session-based Recommendations(CIKM 2018)
• This code is based on pyGRU4REC and original Theano code written by the authors of the GRU4REC paper
• This Version supports TOP1, BPR, TOP1-max, BPR-max, Negative Sampling, Cross-Entropy Losses and other parameters of original Theano code.
• Model achieves the same performance reported in original Theano code example for the Recsys Challenge 2015 (RSC15) dataset.

Requirements

• PyTorch 1.3.1
• Python 3.7
• pandas 0.24.2
• numpy 1.16.2
• jupyter 4.4.0

Usage

Dataset

• Download RecSys Challenge 2015 Dataset from HERE

• Extract Data and locate to PATH_TO_ORIGINAL_DATA directory. Default ( /Data )

Data Preprocessing

For data preprocessing and data cleaning run Preprocessing.py or see Data Preprocessing section in Run_GRU4Rec.ipynb.

• The training set (yoochoose-clicks.dat) itself is divided into training, validation and testing sets where the testing is the last day sessions.
• After preprocessing, rsc15Train.csv, rsc15Valid.csv, rsc15Test.csv obtain that stores in PATH_TO_PROCESSED_DATA directory. Default ( /Data/Cleaned )
• for use only 1/N data before preprocessing change N parameter.

Training And Testing Using Jupyter Notebook

See ``Run_GRU4Rec.ipnyb` that contains:

1. Loads Data

   • Load rsc15Train.csv , rsc15Valid.csv and create Dataset object

2. Model Parameters

   • Set model parameters. in this section parameters set to default like original Theano code.

   • The following list of parameters

   • inputSize GRU Input Size = Number of Items In Dataset
     outputSize GRU output Size = Number of Items In Dataset
     hiddenSize Number of Neurons per GRU Layers (Default = 100)
     nLayers Number of GRU Layers (Default = 1)
     batchSize Mini Batch Size (Default = 32)
     negative Use Negative Sampling In Training Process Or Not (Default = True)
     embeddingDim Size Of The Embedding Used, embeddingDim <= 0 Means Not To Use Embedding (Default = -1)
     dropoutHidden Dropout at each hidden layer (Default = 0.0)
     dropoutEmbed Dropout Of The Input Units, Applicable Only If Embeddings Are Used (Default: 0.0) sigma "Width" Of Initialization. Either The standard Deviation Or The Min/Max Of The Initializations Interval (With Normal And Uniform Initializations Respectively). 0 Means Adaptive Normalization (Sigma Depends On The Size Of The Weight Matrix) (Default: 0.0)
     initAsNormal False: Initializations From Uniform Distribution On [-sigma,sigma]. True: Initializations From Normal Distribution On (0,sigma). (default: False)
     cuda Use GPU Or Not
     finalAct Activation Function (Default = Elu-1.0)
     lossType Type of loss function TOP1 / BPR / TOP1-max / BPR-max / Cross-Entropy/ NLL (Default: BPR-max)
     optimizerType Optimizer (Default = Adagrad)
     lr Learning rate (Default = 0.1).
     weightDecay Weight decay (Default = 0.0)
     momentum Momentum Value (Default = 0.0)
     bpreg Score Regularization Coefficient For The BPR-max Loss Function (Default: 1.0)
     nEpochs Number of epochs (Default = 10)
     timeSort Whether To Ensure The Order Of Sessions Is Chronological (Default: True)
     trainRandomOrder Whether To Randomize The Order Of Sessions In Each Epoch (Default: False)
     sampleAlpha The Probability Of An Item Used As An Additional Negative Sample (Default: 0.75)
     

     trainNSample Number Of Additional Negative Samples To Be Used In Training Mini Batch Generator (Default: 2048)
     

     validNSample Number Of Additional Negative Samples To Be Used In Validation Mini Batch Generator (Default: 2048)
     

     sampleStore Number Of by Precomputing Batch Of Negative Samples (Default: 10000000)
     

     topN Value of K used durig Recall@K and MRR@K Evaluation (Default = 20)
     

3. BPR-max, no embedding

   • train and evaluate the model with best parameters for BPR-max,no embedding

4. BPR-max, constrained embedding

   • train and evaluate the model with best parameters for BPR-max, constrained embedding

5. Cross-entropy

   • train and evaluate the model with best parameters for Log Softmax and Negative Likelihood Loss(Cross Entropy Loss)

6. Testing

   • Load rsc15Test.csv and create test Dataset object
   • Load trained Model and testing