This repo is the TF2.0 implementation of perCLTV: A General System for Personalized Customer Lifetime Value Prediction in Online Games [PDF].

• data/: data of perCLTV (randomly generated sample data to show the data format, not the real data).
  • sample_data_individual_behavior.csv: the sample for individual behavior sequential data.
  • sample_data_social_behavior.csv: the sample for social behavior graph data.
  • sample_data_label.csv: the sample data for label, where label1 is churn label (binary classification) and label2 is payment label (regression).
• src/: implementations of MSDMT.
  • model.py: the code for model.
• main.py: the code for pipeline.

The code has been tested running under Python 3.8.16, with the following packages installed (along with their dependencies):

• tensorflow == 2.12.0
• spektral == 1.2.0
• attention == 5.0.0
• numpy == 1.23.5
• pandas == 2.0.0
• scikit-learn == 1.2.2

Note: to fixup "multiple 'training' args" bug in attention, open attention.py and modify the following code:

  class Attention(Layer):
    
    # ...

    def __call__(self, inputs, training=None, **kwargs):
        if debug_flag:
            # return self.call(inputs, training, **kwargs)
            return self.call(inputs, training=training, **kwargs)
        else:
            # return super(Attention, self).__call__(inputs, training, **kwargs)
            return super(Attention, self).__call__(inputs, training=training, **kwargs)

python main.py 

Please cite our paper if you use this code in your own work:

@article{zhao2023percltv,
  title={perCLTV: A general system for personalized customer lifetime value prediction in online games},
  author={Zhao, Shiwei and Wu, Runze and Tao, Jianrong and Qu, Manhu and Zhao, Minghao and Fan, Changjie and Zhao, Hongke},
  journal={ACM Transactions on Information Systems},
  volume={41},
  number={1},
  pages={1--29},
  year={2023},
  publisher={ACM New York, NY}
}