This is an implementation of the Towards Understanding the Overfitting Phenomenon of Deep Click-Through Rate Models, which is accpete by CIKM2022. The codes are developed with Python2.7 and TensorFlow1.4.

Here, we provide the results on Amazon book data set for reproducibility. In summary, the experiment results of Amazon book data set are consistent with that of production data set. But we recommend the readers to focus on the results of the production data set in the paper, which are more robust to the hyper-parameters and much closer to the online industrial recommender systems.

You can get the data and process it using the script

sh prepare_data.sh

python script/cal_occurrence.py

python script/train.py --model_type DNN --epochs 10 

You can change the default parameters to get the results of different experiments:

• model structure
  • --model_type [DNN,LR]
• corruption percent, which varies from 0.0 (no corruption) to 1.0 (complete random labels).
  • --corruption_percent
• number of parameters
  • --embed_dim
  • --neuron
  • --nlayers
• batch size
  • --batch_size
• activation function
  • --activation [dice,relu,prelu,sigmoid]
• optimizer
  • --optimizer [Adam,sgd,rmsprop]
• techniques to alleviate overfitting
  • --weight_decay
  • --dropout

calculate the A-distance

python script/train_hypothesis.py --model_type DNN --epochs 3

obtain the parameter changes of embedding and MLP layers

python script/train.py --model_type DNN --epochs 3 --print_grad 1