Closed lxx1884896 closed 11 months ago
您好,请教一下,如果我的粗排有多个目标,比如,ctr,cvr, 想问一下在预测时如何进行融合,目前我想到的, 1、 使用multi-head分别对ctr塔和cvr塔的顶部进行提取,将提取得到的ctr embedding以及cvr embedding分别和multi-head提取的user embedding过一遍fe_score函数,然后将 ctr的fe_score和cvr的fe_score 以一定权重进行融合,得到最终的score 2、使用multi-head分别对ctr塔和cvr塔的顶部进行提取,将提取得到的ctr embedding以及cvr embedding以一定权重进行融合,将融合后的embedding 和multi-head提取的user embedding过一遍fe_score函数,得到最终的score 谢谢~
我觉得第一种好一些,你可以通过实验测试一下
您好,请教一下,如果我的粗排有多个目标,比如,ctr,cvr, 想问一下在预测时如何进行融合,目前我想到的, 1、 使用multi-head分别对ctr塔和cvr塔的顶部进行提取,将提取得到的ctr embedding以及cvr embedding分别和multi-head提取的user embedding过一遍fe_score函数,然后将 ctr的fe_score和cvr的fe_score 以一定权重进行融合,得到最终的score 2、使用multi-head分别对ctr塔和cvr塔的顶部进行提取,将提取得到的ctr embedding以及cvr embedding以一定权重进行融合,将融合后的embedding 和multi-head提取的user embedding过一遍fe_score函数,得到最终的score 谢谢~