Statement

this is the official code for sigir 2022 paper 《CTnoCVR: A Novelty Auxiliary Task Making the Lower-CTR-Higher-CVR Upper》

paper is here https://dl.acm.org/doi/pdf/10.1145/3477495.3531843

structure of the code

• the main code is in mctr
• scripts for running is in tools
• put as much as possible parameters in a yaml file in configs, which makes it easier to manage and also change in commandline.

how to run

• all parameters should be in a yaml file as shown in configs; you have to change location of training/eval files
• tools/train.py control the workflow.
• python train.py -c configs/esmm_ali_cpp.yml
• you can override parameters as follows:
  • python esmm_model.py -c configs/esmm_ali_cpp.yml -o Global.epochs=4 Models.esmm.mlp_ctr="[360, 200, 80]"
  • more examples
    • CUDA_VISIBLE_DEVICES=2 python tools/train.py -c configs/esmm_ali_cpp.yml -o Loss.weights="[1., 1., 1.]" Loss.value="[direct_auc_loss, direct_auc_loss, fake_loss]"

TODO

• figure out how to utilize the value of each feature (currently didn't pay attention to it)
• make datareader faster (currently 350ms/batch(5000 example))
• all negative batch could impact the learning process since it's easy to overfit due to the imbalance of ali-cpp dataset. I guess we could try to give a dynamic weight of each batch according to the pecentage of negative examples in a batch. added focalLoss, but it's only comparable with binary_crossentropy loss.

contributor

jeffzhengye scalaboy