应用背景:online food recommendation service (OFRS)
· 时空特征不是越多越好,精细化筛选可以提升模型效果;
· 提出Dynamic Context Adaptation Model
Baseline Model
四类特征(user + item + behavior + context) + multi-head target attention, 目标是CTR;
通过人工特征筛选,AUC提升了1.45pt
Sequence Modeling
time difference指:We incorporated the time difference between the sequence time and the target time as a feature in the sequence so that the model could learn the temporal orderliness.
DYNAMIC CONTEXT ADAPTATION MODEL
模型结构:
增加一个动态特征选择的模块来挑选特征,提升效果,减轻人工筛选的工作;
k的选择对auc的影响:𝑘 is set to 4 by grid search