Implements of Awesome RecSystem Models with PyTorch/TF2.0
The paper is available at: https://arxiv.org/pdf/1708.05123.pdf
Tested dataset: Criteo
Run data/forDCN/DCN_dataPreprocess_PyTorch.py to pre-process the data. According to the paper, the data set is split by 9:0.5:0.5 for train, test and valid
Split the data set by 9:0.5:0.5 for train, test and valid
Run Model/DeepCrossNetwork_PyTorch.py, and the results are as follows:
Epochs | AUC | LogLoss |
---|---|---|
1st | 0.80157 | 0.45192 |
2nd | 0.80430 | 0.44922 |
3rd | 0.80546 | 0.44817 |
4th | 0.80639 | 0.44729 |
5th | 0.80696 | 0.44678 |
Tested dataset: Criteo
Split the data set by 9:1 for train and test.
How To Run: Run data/forOtherModels/dataPreprocess_PyTorch.py to pre-process the data, then run Model/ProductNeuralNetwork_PyTorch.py
After 5 epochs, the results are AUC:
Framework(Algorithm) | AUC | LogLoss |
---|---|---|
PyTorch(IPNN) | 0.76585 | 0.47766 |
PyTorch(OPNN) | 0.77656 | 0.46988 |
TensorFlow(IPNN) | 0.77996 | 0.46827 |
TensorFlow(OPNN) | 0.78098 | 0.46718 |