Many applications typically require targeted display ads for customer acquisition. For credit card advertising, consistent and effective customer acquisition is more challenging than traditional advertising due to the long link between user conversion. Based on the practice of Meituan’s co-named credit card business and the paper published in KDD 2021 this year, this paper introduces an adaptive information Migration multi-task (AITM) framework, which can model the sequential dependency between users’ multi-step transformations and improve the high-end end-to-end customer acquisition conversion rate. I hope it can be helpful or enlightening to students engaged in related research.

Thesis download: Modeling the Sequential Dependence among Audience multi-step Conversions with multi-task Learning in Targeted Display Advertising”

Source code: github.com/xidongbo/AI…

A, in this paper,

Many apps, such as e-commerce platforms or finance apps, often need targeted display ads to reach customers. With the explosive growth of these applications, one of the biggest challenges for these large-scale applications in the real world is how to achieve consistent and effective customer acquisition. In these applications, customer acquisition is usually a multi-step user conversion process. For example, in e-commerce platforms, user conversion is usually a process of exposure -> click -> purchase. In financial advertising, such as credit card business, user conversion is usually a process of exposure -> click -> apply -> check card -> activation. So, getting customers in credit card advertising is more challenging than traditional advertising:

  • The dependency of path sequence in user multistep transformation is longer.
  • In the long sequence dependence of multi-step transformation, the positive feedback (positive sample) is gradually sparse, and the category imbalance becomes more and more serious.

In meituan’s credit card business, we usually expect the user to complete the last two conversion stages (i.e., exposure -> core card, and exposure -> activation) before the conversion is considered effective. Therefore, it is necessary to use the rich positive sample information of the front task to alleviate the category imbalance of the follow-up task. In this direction, multi-tasking learning is a typical solution to improve the end-to-end customer acquisition conversion rate.

Second, background introduction

In our credit card business:

  • Impression: It means that the AD is presented to a specific user, who is selected according to some sort of index, such as CTR, CVR, etc.
  • Click: If the user is interested in the content of the AD, he will Click on the AD. After clicking, he will be taken to the application form page.
  • Application: After entering the Application form page, users can fill in the Application form and click the “Apply” button to apply for a credit card.
  • Approval: Also known as Approval, which means that a user has good credit and has applied for and been granted a certain credit limit. In our system, this is also a process of systematic real-time judgment in most cases.
  • Activation: After the credit is granted and the credit card is mailed, the user can activate the credit card and use it. Active or not tags are often difficult to obtain because of the time it takes for credit cards to be mailed and the performance period required for users to actively deactivate them, so the category imbalance is more severe here. In this case, we usually look to see if the user will activate the credit card within 14 days of checking it.

The above transformation steps have sequence dependencies, which means that only if the previous step occurs can the next step occur. Based on this constraint, users can only have five different transformation stages, as shown in the figure below:

In industry and academia, multi-tasking learning is a typical solution for increasing end-to-end customer acquisition conversion. In recent years, great progress has been made in modeling the relationship between tasks in multi-task learning. We divide these major studies into two broad categories:

1) Control how the Expert module at the Bottom of the multi-task model shares parameters between tasks [1,2,3], and the multi-tower module at the top handles each task separately, as shown in Figure 3 (a). We call this category Expert-Bottom ** mode. However, the expert base mode can only transmit shallow representations between tasks, but tends to contain richer and more useful representations in the network near the output layer [4,5], which has been shown to yield more gains [6]. In addition, because expert base patterns are not specifically designed for tasks with sequence dependencies, these models with expert base patterns cannot explicitly model sequence dependencies.

2) migration Probability [7,8,9,10] in the output layer of different tasks, as shown in FIG. 3 (b), which is called the ** Probability Transfer ** mode. The probability transfer mode can only transfer the simple probability information through scalar product, but ignores the richer and more useful representation in vector space, resulting in the loss of information. If any one of these probabilities is not accurately predicted, multiple tasks will suffer.

In this paper, an Adaptive Information Transfer multi-task (AITM) framework is proposed for sequence-dependent tasks. The AIT module is used to model the sequential dependencies between users’ multi-step transformations. The AIT module adaptively learns what and how much information needs to be transferred at different transformation stages. In addition, the AITM framework can more accurately identify end-to-end transformations by including a behavior expectation calibrator in the loss function. The framework is deployed in the Meituan App to display ads for meituan co-branded cards in real time for users with high conversion rates.

Iii. System overview

The figure above shows a schematic of a multitasking sorting system. In our credit card business, we modeled four tasks in addition to the passive exposure step. The core card, activation is the main task, click, application is the auxiliary task. This is because if the user only completes the click and apply step, and the card check step is not complete, then it is not a valid transformation and will waste resources (such as computing and traffic resources). Therefore, we focus on the last two end-to-end transformation tasks, namely exposure -> core card and exposure -> activation. Because the number of positive samples in the latter two tasks is small and the activation feedback is delayed, while the number of positive samples in the first two auxiliary tasks is large, the problem of category imbalance in the latter two tasks can be alleviated by the first two tasks.

In addition, Meituan co-branded credit cards are issued in cooperation with different banks, which are in different stages of business development, so they have different requirements for card verification and activation. Banks that are just starting out in the credit card business often want to issue more cards to capture the market quickly, while mature banks want to increase activation rates to make a quick profit. Therefore, we have a selector in our system that outputs different transformation targets for different banks, and a multi-tasking framework that handles different business requirements well.

In addition, because different businesses of Meituan need traffic to obtain customers for their respective businesses, and different users have different interests in different businesses, it is not possible to simply allocate traffic to different businesses. We need a multi-service sorting mechanism to allocate traffic to maximize the overall benefit.

Iv. Model introduction

Figure (c) above shows our proposed AITM framework. The framework uses the AIT module to model the sequence dependencies in the user’s multi-step transformation process. This model diagram shows two adjacent tasks: T −1t-1t−1 and TTT.

Firstly, the input feature x\ PMB {x}xx is output to multiple Tower networks respectively through the task sharing Embedding. Through the shared Embedding module, the Embedding representation can be learned by using the rich positive sample information of the preceding tasks, so as to alleviate the problem of category imbalance in subsequent tasks; on the other hand, the model parameters can be reduced. The Tower network can be customized according to your own needs, here we only used a simple MLP. Then, the AIT module uses the vector QT \ PMB {q} _tQQt output from the Tower of the current task and the information pt−1\ PMB {p}_{T-1} PPT −1 from the previous task to learn how to fuse information between tasks. The AIT module is defined as follows:


z t = A I T ( p t – 1 . q t ) . \pmb{z}_t=AIT(\pmb{p}_{t-1},\pmb{q}_t),

p t – 1 = g t – 1 ( z t – 1 ) . \pmb{p}_{t-1}=g_{t-1}(\pmb{z}_{t-1}).

In fact, a specially designed attention mechanism is used to automatically assign weights to the migration information PT −1\ PMB {p}_{t-1} PPT −1 and the original information QT \ PMB {q} _tQQt. And transfer the information of pt – 1 \ PMB {p} _ {} t – 1 PPT – 1 is a function by gt – 1 (⋅) g_ {1} t – (,) gt – 1 (⋅) to study, here gt – 1 (⋅) g_ {1} t – (,) gt – 1 (⋅) can be a simple full connection layer, To learn what information to migrate between two adjacent tasks. The input zt−1\ PMB {z}_{t-1} ZZT −1 is the output of the AIT module of the previous task.

Specifically, the attention mechanism in the AIT module is designed as follows:


z t = ∑ u ∈ { p t – 1 . q t } w u h 1 ( u ) . \pmb{z}_t=\sum_{\pmb{u}\in\{\pmb{p}_{t-1},\pmb{q}_t\}}w_uh_1(\pmb{u}),

w u = e x p ( w ^ u ) ∑ u e x p ( w ^ u ) .   w ^ u = < h 2 ( u ) . h 3 ( u ) > k . w_u=\frac{exp(\hat{w}_u)}{\sum_u exp(\hat{w}_u)},~\hat{w}_u=\frac{< h_2(\pmb{u}),h_3(\pmb{u})>}{\sqrt{k}}.

Where, wuW_uwu is the weight of automatically learned transfer information. H1 (⋅) h_1 (,) h1 (⋅) and h2 (⋅) h_2, (), h2 (⋅), h3 (⋅) h_3 (,) h3 (⋅) feedforward neural network input projection to a new vector space.

Finally, we constrain the output of the probability to satisfy the sequence dependence as much as possible by applying a calibrator in the loss function. The loss function is the cross-entropy + calibrator constraint, where α∈[0,1]\alpha\in[0,1]α∈[0,1] defines the calibrator constraint strength:


L ( Theta. ) = L c e ( Theta. ) + Alpha. L l c ( Theta. ) . L(\theta)=L_{ce}(\theta)+\alpha L_{lc}(\theta).

Specifically, cross entropy is defined as follows:


L c e ( Theta. ) = – 1 N ∑ t = 1 T ∑ ( x . y t ) ∈ D N ( y t l o g y ^ t + ( 1 – y t ) l o g ( 1 – y ^ t ) ) . L_{ce}(\theta)=-\frac{1}{N}\sum_{t=1}^T\sum_{(\pmb{x},y_t)\in D}^N \left(y_tlog\hat{y}_t+(1-y_t)log(1-\hat{y}_t)\right).

The calibrator constraints are defined as follows:


L l c ( Theta. ) = 1 N ∑ t = 2 T ∑ x ∈ D N m a x ( y ^ t – y ^ t – 1 . 0 ) . L_{lc}(\theta)=\frac{1}{N}\sum_{t=2}^T\sum_{\pmb{x}\in D}^Nmax(\hat{y}_t-\hat{y}_{t-1},0).

If y^t\hat{y}_ty^t is greater than y^t−1\hat{y}_{t-1}y^t−1, the calibrator will output a positive penalty term, otherwise 0.

V. Experimental analysis

In this section, we will evaluate our proposed AITM framework by conducting experiments on industrial and publicly available real-world data sets against various baseline models. Next, we will first introduce the data set used, then the results of the offline and online experiments, and finally further experimental analysis (including ablation experiments, hyperparametric experiments, and Case studies).

5.1 the data set

We conducted an offline experiment using two data sets. The industrial data set is all the samples of advertising exposure of Meituan co-branded cards, and we took samples over a period of time. The dataset has four tasks: click, apply, check the card, and activate. In our business, we only focus on the core card and the activation of the conversion metrics, the two steps of transformation is effective transformation. The public data set is the click-transform estimate data set of ali [9,11], which has two tasks: click and buy.

5.2 Offline and online experiments

We did both offline and online experiments. With the development of the business, we have successively deployed LightGBM, MLP and AITM models online. Both offline and online experiments have demonstrated the remarkable performance of AITM.

5.3 Ablation experiment

We designed the AIT module to model the sequential dependencies in the user’s multi-step transformation process. To demonstrate the effectiveness of the AIT module, we conducted ablation studies on the AIT module. We randomly selected 500 positive and negative activation samples, and then sorted their activation task prediction scores. The higher the predicted score, the more likely the model predicts that these users will activate the co-branded card. Then we use T-SNE to draw a two-dimensional diagram of the original information qt\ PMB {q} _tQQt, migration + original information PT −1\ PMB {p}_{t-1} PPT −1+ qt\ PMB {q}_tqqt, and the information zt\ PMB {z}_tzzt learned by AIT module.

From the figure below, we can see that when user activation is relatively easy to predict (i.e., activation score top0% – top50%), all three modules can distinguish positive and negative samples well. However, when the user activation is more difficult to predict (i.e., the activation score is top50% – top100%), the original information qt\ PMB {q} _tQQt, migration + original information pt−1\ PMB {p}_{t-1} PPT −1+ qt\ PMB {q} _tQQt can not well distinguish the positive and negative samples. The AIT module is significantly better than the other two parts in the top50% – top100% prediction score, which indicates the effectiveness of the AIT module.

5.4 Hyperparameter experiment

In addition, in order to study the stability of the AITM framework and its dependence on the hyperparameters, we conducted a hyperparameter study. By taking different values of Embedding vector dimension, calibrator strength, positive sample proportion and task number, we conducted several experiments. We found the following findings:

  • The first two hyperparameter experiments show that the AITM module is stable under the hyperparameters of Embedding vector dimension and calibrator strength.
  • The third hyperparameter experiment shows that an appropriate downsampling ratio is an important measure to alleviate the category imbalance.
  • The final figure shows how more tasks can lead to more useful information, further improving the end-to-end conversion rate.

5.5 Case Study

Finally, we conducted a Case Study to understand what and how much information the AIT module migrated at different stages. WuW_uWu in the figure below is the weight of the migration information. We first randomly sampled 40,000 test samples, then divided them into three groups according to the Label of the two adjacent tasks: 00/10/11 (corresponding to the red, blue and green lines in the figure respectively), and ranked the Top 500 samples in each group according to Logloss. The higher the Top sample, the more accurate the prediction.

  • First, look at the red line, because of the sequence dependency, if the current task Label is 0, the next task Label can only be 0. So we see that the former task migrates very strong information to the latter task (weight is almost 1).
  • Then, looking at the blue and green lines, if the current task Label is 1, the latter task Label is uncertain. As Logloss increases, the weight of the migration information increases, suggesting that the prediction results of the latter task may be misleading by the previous task.
  • In addition, if you look at the green line, when the Label of the latter task is 1, the former task migrates very little information (the weight is very small), which indicates that the latter task mainly identifies positive samples according to its own task.
  • From the above results, we can see that the AIT module can learn how much information should be transferred between two adjacent tasks.

Six, summarized

In this paper, an adaptive information transfer multitasking (AITM) framework is proposed to model the sequential dependencies between user multistep transformations. The proposed adaptive information transfer (AIT) module, combined with the behavior expectation calibrator, can learn what and how much information needs to be transferred in different transformation stages to improve the performance of sequential dependent multi-task learning. The results of offline and online experiments show that the framework is a significant improvement over the latest baseline model. The framework is deployed in the Meituan App to display ads for meituan co-branded cards in real time for users with high conversion rates. For more technical details, please refer to our paper.

Vii. Introduction to the author

  • Dong Bo, Chen Zhen, Yan Peng, Chen Yu and others are from Meituan Financial Intelligence Application team
  • Zhang Yingerer: Intern of Meituan, undergraduate of Zhejiang University
  • Zhu Yongchun: PhD student at the Institute of Computing Science, Chinese Academy of Sciences
  • Zhuang Fuzhen: Researcher at the Institute of Artificial Intelligence, Beihang University

Recruitment information

The algorithm position of Meituan Financial Intelligence application team continues to attract excellent algorithm engineers and experts, coordinate in Beijing/Shanghai. Available positions include:

Marketing algorithm engineer/expert

  • Serving various business scenarios of Meituan Finance, responsible for the algorithm design and development of marketing customer acquisition, retention and promotion, and integrating machine learning and optimization techniques to solve financial marketing problems;
  • Precipitation algorithm platform capability, improve the efficiency of algorithm application, provide customer group mining, equity allocation, material matching, dynamic creativity, operations planning, precision reach and other intelligent solutions;
  • In combination with the financial business scenario of Meituan, it explores and innovates the cutting-edge ARTIFICIAL intelligence technologies such as deep learning, reinforcement learning and knowledge graph, and implements the precipitation and implementation of innovative technologies.

Risk control algorithm engineer/expert

  • Through the development and optimization of machine learning model and strategy, continuously improve the ability to identify financial risk behaviors;
  • In-depth understanding of the business, the application of machine learning technology to improve the automation of risk control work, comprehensively improve business efficiency;
  • Follow up the cutting-edge technology of artificial intelligence, and explore the landing in the financial risk control scenario.

NLP algorithm engineer/expert

  • Based on the financial business scenarios of Meituan, combined with natural language processing and machine learning technologies, the intelligent dialogue robot is deployed to financial marketing, risk management, customer service and other scenarios.
  • Participate in the research and development of dialogue robot related projects, including but not limited to semantic understanding, multi-round dialogue management and other related algorithm development and optimization;
  • Continuously follow up the development of relevant technology in academia and industry, and quickly apply it to the project.

Interested students are welcome to send their resume to [email protected] (subject line: Meituan Financial Intelligence Application Team).

reference

  • [1] Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H Chi. 2018. Modeling task relationships in multi-task learning with multi-gate mixture-of-experts. In KDD. 1930–1939.
  • [2] Zhen Qin, Yicheng Cheng, Zhe Zhao, Zhe Chen, Donald Metzler, And Jingzheng Qin. 2020. Multitask Mixture of Sequential Experts for User Activity Streams. In KDD. 3083 — 3091.
  • [3] Hongyan Tang, Junning Liu, Ming Zhao, and Xudong Gong. 2020. Progressive Layered Extraction (PLE): A Novel Multi-task Learning (MTL) Model for Personalized Recommendations. In RecSys. 269 — 278.
  • [4] Yixuan Li, Jason Yosinski, Jeff Clune, Hod Lipson, and John E Hopcroft. 2016. Convergent Learning: Do different neural networks learn the same representations? In ICLR.
  • [5] Matthew D Zeiler and Rob Fergus. 2014. Visualizing and understanding convolutional networks.
  • [6] Eric Tzeng, Judy Hoffman, Ning Zhang, Kate Saenko, and Trevor Darrell. 2014. Deep domain confusion: ArXiv Preprint arXiv:1412.3474 (2014).
  • [7] Chen Gao, Xiangnan He, Dahua Gan, Xiangning Chen, Fuli Feng, Yong Li, Tat-Seng Chua, And Depeng Jin-2019. Neural Multi-task Recommendation from multi-behavior data. In ICDE.1554 — 1557.
  • [8] Chen Gao, Xiangnan He, Danhua Gan, Xiangning Chen, Fuli Feng, Yong Li, Tat-Seng Chua, Lina Yao, Yang Song, and Depeng Jin. 2019. Learning to Recommend with Multiple Cascading Behaviors. TKDE (2019).
  • [9] Xiao Ma, Liqin Zhao, Guan Huang, ZhiWang, Zelin Hu, Xiaoqiang Zhu, and Kun Gai. 2018. Entire space multi-task model: An effective approach for Estimating post-click conversion rate. In SIGIR. 1137 — 1140.
  • [10] Hong Wen, Jing Zhang, Yuan Wang, Fuyu Lv, Wentian Bao, Quan Lin, and Keping Yang. 2020. Entire Space Multi-Task Modeling via Post-Click Behavior Decomposition for Conversion Rate Prediction. In SIGIR who lead. 2377-2386.
  • [11] tianchi.aliyun.com/datalab/dat…

Read more technical articles from the Meituan technical team

Front end | | algorithm back-end | | | data security operations | iOS | Android | test

| in the public bar menu dialog reply goodies for [2020], [2019] special purchases, goodies for [2018], [2017] special purchases such as keywords, to view Meituan technology team calendar year essay collection.

| this paper Meituan produced by the technical team, the copyright ownership Meituan. You are welcome to reprint or use the content of this article for non-commercial purposes such as sharing and communication. Please note “Reprinted from Meituan technical team”. This article may not be reproduced or used commercially without permission. For any commercial activities, please send an email to [email protected] to apply for authorization.