Esmm reading notes

The paper ：《Entire Space Multi-Task Model: An Effective Approach for Estimating Post-Click Conversion Rate》 Ali ,2018

1.motivation

differ CTR Estimate the problem ,CVR There are two key issues in the estimation ：

Sample Selection Bias (SSB)  Conversion is after clicking “ There may be ” What happened , Tradition CVR The model usually takes click data as training set , The click is not converted to a negative example , Click and convert to positive example . But when the model predicts , It's an estimate of the entire space , Instead of just estimating the click samples . That is , The training data and the actual data to be predicted come from different distributions , The training data is the data of the click Set , The forecast data is full data . Use a finite subset of data to predict the entire sample , There will be large deviations . To see only one spot , It can be seen .

chart 1 SSB Sketch Map  

Data Sparsity (DS)  As CVR The click sample of training data is far less than CTR Estimate exposure samples used for training . Data sparsity is fitting cvr A difficulty in modeling .

Some strategies can alleviate these two problems , For example, from the exposure set to unclicked Sample sampling for negative case mitigation SSB, Oversampling of transformed samples DS etc. . But either way , Not very elegant Effectively solve any of the above problems .

You can see ： Click on —> conversion , Itself is two strongly related continuous behaviors , The author hopes to show that this kind of “ Behavior chain relationship ”, Thus, training and prediction can be carried out in the whole space . This involves CTR And CVR Two tasks , So use multitasking （MTL） It's a natural choice , The key highlight of the paper is “ How to build ” This MTL.

2.model

From the perspective of modeling SSB and DS problem .

First, clarify three concepts ：

CTR : Exposed , Click probability

CVR ： If item By clicking the , So the probability of its transformation . Pay attention to the assumptions “ If it's clicked ”

CTCVR  ：item By clicking the , The probability of its transformation

It is not possible to train directly with all samples CVR The reason for the model is ：

I do not know! unclicked Of item, Suppose they were user Click. , Whether they will be transformed . If used directly 0 As their label, Will be largely misleading CVR Model learning .

Can't get “unclicked Of item, Suppose they were user Click. ” This information , Give to the CVR Model fitting in progress .

  among y, z respectively conversion and click. be aware , In all sample spaces ,CTR Corresponding label by click, and CTCVR Corresponding label by click & conversion

  chart 2. ESMM Network structure

cvr The model is built on all sample spaces , That is, you can use all the data to fit cvr Model ;ctr and cvr The lower layer of the network is

Take a closer look at the picture above , Pay attention to the following points ：1） share Embedding CVR-task and CTR-task Use the same features and features embedding, That is, both from Concatenate After that, I will learn the exclusive parameters of each part ;2） Implicit learning pCVR  What do you mean ？ here pCVR（ Pink node ） Just one of the networks variable, There is no monitoring signal shown .

  It is using CTCVR and CTR Monitoring information to train the network , Learn implicitly CVR, That's exactly what it is. ESMM The essence of , As for the necessity and rationality of doing so

Think again ,ESMM The structure of is based on “ ride ” Relationship design ——pCTCVR=pCVR*pCTR, Is it possible to pass “ except ” Our relationship has been pCVR, namely pCVR = pCTCVR / pCTR ？ For example, train one CTCVR and CTR Model , Then divide and get pCVR, In fact, it can be , But there is an obvious drawback ： Predicted by the real scene pCTR、pCTCVR The values are relatively small ,“ except ” It is easy to cause numerical instability . The author compares this method in the experiment .