Skip to content

JianzhouZhan/Awesome-RecSystem-Models

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

78 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Implements of Awesome RecSystem Models with PyTorch/TF2.0

1. Requirements

  • TensorFlow2.0, PyTorch1.2+, Python3.6, NumPy, sk-learn, Pandas

2. DataSets

  • Criteo
    • This dataset Contains about 45 million records. There are 13 features taking integer values (mostly count features) and 26 categorical features.
    • The columns are tab separated with the following sechema: <int feat 1>...<int feat 13><cate feat 1>...
    • The dataset is available at http://labs.criteo.com/2014/02/download-kaggle-display-advertising-challenge-dataset/
    • Put the downloaded 'train.txt' file in 'data/Criteo/' folder.
  • Movielens100K
    • MovieLens 100K movie ratings. Stable benchmark dataset for Recommendation System.
    • This dataset contain 100,000 ratings from 1000 users on 1700 movies.
    • The details can be found at https://grouplens.org/datasets/movielens/100k/
    • This dataset have been downloaded and is available at 'data/Movielens100K'

3. Implemented Models:

  • FM: Factorization Machine
    • The paper is available at: https://www.csie.ntu.edu.tw/~b97053/paper/Rendle2010FM.pdf
    • 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/FM_PyTorch.py. Or run data/forOtherModels/dataPreprocess_TensorFlow.py to pre-process the data, and run Model/FM_TensorFlow.py
    • the result is AUC: 0.7805(PyTorch), 0.7791(TensorFlow)
    • For Multi-Classification implements:
      • Tested dataset: Movielens100K
      • Run FM_Multi_PyTorch.py
  • FFM: Field-aware Factorization Machine
  • DeepFM: Factorization-Machine based Neural Network
    • The paper is available at: https://www.ijcai.org/proceedings/2017/0239.pdf
    • 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/DeepFM_PyTorch.py. Or run data/forOtherModels/dataPreprocess_TensorFlow.py to pre-process the data, and run Model/DeepFM_TensorFlow.py
    • PyTorch: After 3 Epochs, AUC: 0.795(The paper result is 0.801)
    • TensorFlow: After 3 Epochs, AUC: 0.8014, LogLoss: 0.4516
  • DCN: Deep&Cross Network
    • 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
  • xDeepFM: eXtreme Deep Factorization Machine
    • The Paper is available at: https://arxiv.org/pdf/1803.05170.pdf
    • Tested dataset: Criteo
    • Split the data set by 9:1 for train and test.
    • Run data/forXDeepFM/xDeepFM_dataPreprocess_PyTorch.py to pre-process the data. The pre-process of xDeepFM is identical with that of DeepFM.
    • Run Model/xDeepFM_PyTorch.py.
    • After 5 epochs, the result is AUC 0.80148, LogLoss 0.45104 (The paper result is AUC 0.8052).
  • PNN: Product-based Neural Network
    • Paper: https://arxiv.org/pdf/1611.00144.pdf

    • 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

Releases

No releases published

Packages

No packages published

Languages