ESMM

Advantage: Combine two nn together with by a dot product finally. Key word: DNN Usage: Predict CVR and CTR

Predicting post-click conversion rate is the problem they try to solve.

Why the Problem Is Difficult

Current method has two problems, one is illustrated as the following.

1. Sample selection bias problem means that the training space is smaller than the inference space, which makes the inference inaccurate.

   

2. Data Sparsity Problem Data for CVR problems is much less than CTR task.

Problem Formulation

Post-click CVR estimation could be represented as predicting the following value:

\[p(z=1|y=1,x)\]

Post-view click&conversion rate’s relationship with pCTR & pCVR. Product.

传统做CVR预测的时候都使用了一个辅助向量空间，这个space中的数据具有一个特性，就是以下，对于这个sub-space里面的所有feature vector，他们都是被clicked的。所以一个预测的方法难道不是先预测哪些可以被点击，在预测点击后哪些可能被转化吗？

Solution

从problem formulation这个角度来讲，从最上面的那个公式做了个转化，

直接求公式右上的，就是一个乘积的形式，避免了除法带来的风险。

Loss Function：

l是cross-entropy loss。