This is the code of the paper Solving the Loss Imbalance Problem in Live Streaming Recommendation with Gradient Normalization
Due to business sensitive issues, our data cannot be shared, so only the code and feature logic are shared.
The comparison method in the paper can be implemented as follows:
ESMM:
python esmm_train.py --conf esmm_train_conf.ini
ESMM+GradNorm:
python esmm_train_gn.py --conf esmm_train_conf_gn.ini
ESMM+FM:
python esmm_train_fm.py --conf esmm_train_conf_fm.ini
ESMM+FM+GradNorm:
python esmm_train_fm_gn.py --conf esmm_train_conf_fm_gn.ini
ESMM+FM+MMOE:
python esmm_train_mmoe_fm.py --conf esmm_train_conf_mmoe_fm.ini
ESMM+FM+MMOE+GradNorm:
python esmm_train_mmoe_fm_gn.py --conf esmm_train_conf_mmoe_fm_gn.ini