[recommendation system] breakthrough and imagination of deep location interactive network dpin for meituan takeout recommendation scenario

Today, I'd like to introduce an article about meituan in SIGIR 2021 A Chinese manuscript paper, This paper mainly introduces how to solve the problem of position offset in hit rate estimation （position-bias） Some work on , Let's take a look .

Meituan basic R & D machine learning platform training engine team , Unite with the algorithm efficiency team of home search and push technology department 、NVIDIA
DevTech The team , Set up a joint project team . At present, it has been deployed in the meituan takeout recommendation scenario , Offline effect of multi generation model full alignment algorithm , Before comparison , The optimized CPU Mission , The cost performance has improved 2~4 times .

1、 background

Click through rate (CTR) Forecasting plays an important role in online advertising and recommendation systems . In practice , Yes CTR Model training depends on click data , And in a higher position , In essence, it tends to a higher position , Because the higher position has a higher CTR. Existing methods , Such as actual position training , It has fixed position reasoning and inverse tendency weighted training , Positional reasoning , The deviation problem is reduced .
In the past, some work has been done to solve the problem of position offset . The most common approach is to take the position feature as a feature of a model training , And online prediction , All candidate advertisements use the same location feature input . The implementation of this scheme is relatively simple , But online prediction , Choose a different location , There will be differences in the recommended results , The result is often suboptimal .

Huawei put forward PAL The framework divides the probability of an advertisement being clicked into two factors ： The probability that the advertisement is seen by the user and the probability that the user clicks after seeing the advertisement . The paper makes further assumptions ： Whether users see advertisements is only related to the location of advertisements ; meanwhile , After the user sees the advertisement , Whether to click on the advertisement has nothing to do with the location . Therefore, the whole framework also includes two parts , As shown in the figure below . When forecasting online , Just deploy the network on the right , The hit rate obtained is the hit rate after eliminating the position offset . The disadvantage of this scheme is that the assumptions are too strong , Oversimplify the problem , Position offset and user characteristics are not fully considered 、 Contextual features and candidates item The relationship between .

2、DPIN Introduce

DPIN It consists of three modules , They are dealing with 𝐽 Basic module of item 、 Handle 𝐾 The depth of items interacts with modules and combinations by location 𝐽 Xiang He 𝐾 Item's positional grouping module . Through the network , It is possible to predict the performance of all candidates for each position under the constraints of service performance CTRs.

2.1 base module

With most depths CTR The model is similar , We use an embedded structure ,MLP (Multiple Layer Perception) As our basic module . For a request , The basic module will be a user 、 A context and 𝐽 Candidate entries as input , Then use the request information 𝒓 It means the first one j Entries ：

2.2 deep position-wise interaction module

The module independently retrieves the sequence of user behavior at each location , Conduct location-based interest aggregation , Eliminate the position deviation of lifelong sequence . then , Use location 、 Positional nonlinear interaction between environment and user . Last , The transformer is used to realize the deep interaction between different positions .

2.3 position-wise combination module

The purpose of the position combination module is to combine J Items and predict the click through rate of each item in each location K The location of .𝒓 Nonlinear interaction between 𝒌 and 𝐸(𝑘) For learning users 、 Context 、 The nonlinear relationship between items and positions . The first k The third position is j Term CTR by 𝐶𝑇𝑅𝑗𝑘 The calculation method of is as follows :

3、 Experimental results and Analysis

Finally, let's look at the experimental results ：
The following table shows the experimental results of the comparison method on conventional and random test sets . We first analyze the differences of different methods in conventional traffic . And DIN comparison ,DIN+PosInWide and DIN+PAL Method in AUC The performance of has declined , But in PAUC Improved in , This shows that both methods can effectively alleviate the position deviation , But it can lead to inconsistencies between offline and online .

The service delay and... Of the bit difference combination module DIN The model is insignificant , Because the user sequence operation has a large delay .DPIN The service delay of is slowly increasing , Because the deep interaction module has nothing to do with the project . And DIPIN + ItemAction comparison ,DPIN The service performance has been greatly improved , No damage to model performance , This shows that our method is both effective and efficient .

Okay , That's all for this article ～