Exploration and practice of heterogeneous advertising in meituan in-store business

Under the restriction of LBS (Location Based Services) distance, fewer candidates restrict the potential space of the whole in-store advertising ranking system. In this paper, we introduce the candidate extension from the perspective of candidate types, and solve the performance challenge by using high-performance heterogeneous hybrid network, thus raising the upper limit of the potential of local life scene ranking system. Hope to be engaged in the relevant direction of the students to inspire.

1 Background and introduction

1.1 background

Meituan in-store advertising is responsible for the commercial realization of Meituan search traffic, serving in-store catering, family entertainment, beauty medical beauty, hotel tourism and many other local life service businesses. The quality estimation team is responsible for CTR/CVR and customer unit price/transaction volume estimation in advertising systems. In the past few years, we have made some breakthroughs in user and context estimation through innovative technologies such as rank context modeling [1] and ultra-long temporal sequence modeling [2] [3]. The paper was published in SIGIR, ICDE, CIKM and other international conferences.

However, the above paper focuses on model accuracy, which together with advertising candidates determines the quality of the ranking system. However, from the perspective of advertising candidates, compared with the candidate set of traditional e-commerce, meituan search advertising has fewer store candidates in some categories due to the restriction of LBS (Location Based Services), which seriously restricts the potential space of the whole sorting system. When the traditional method of increasing the number of candidates failed to achieve profits, we considered to expand and optimize the advertising candidates in order to raise the maximum potential of the local life scene ranking system.

1.2 Scenario Introduction

A single store advertisement is not enough to meet the fine-grained intention demands of users looking for goods and services. In some scenarios, commodity advertisement is used as the candidate supplement of store advertisement. In addition, there are some scenes of commodity advertisements hanging in the form of combination with store advertising display. Various forms of heterogeneous advertising display styles have brought opportunities and challenges to the in-store advertising technology team. According to the characteristics of business scenes, we have optimized the mixed arrangement of heterogeneous advertising. Taking the marriage channel page of Meituan and the homepage search of Meituan as examples, the following sections respectively introduce two types of typical heterogeneous mixed advertising: competitive heterogeneous advertising and combinational heterogeneous advertising.

Heterogeneous advertising: store and commodity advertising are competitive and mixed, and the display type of advertising is determined by comparing pCTR in the mixed model. As shown in Figure 1 below, the first place in the left column is the store type advertisement, showing store pictures, store titles and the number of star reviews; The first item in the right column is the product type advertisement, showing the product picture, product title and corresponding stores. The advertising system determines the order of the ads and the type of display, and when the commodity type AD wins, the system determines the specific product to be displayed.

Combination relation Heterogeneous advertisement: store advertisement and its product advertisement combination are listed and sorted as a display unit (blue box). Products are subordinate to the store, and the two types of heterogeneous advertisement combination are displayed in a mixed row. As shown in Figure 2 below, store advertisements display information such as the head picture and title price of stores; Two product ads display information such as price, title and sales volume. The advertising system determines the order of display units and the Top2 items to be displayed in the store’s collection.

1.3 Introduction to challenges and Practices

At present, online search advertising model is a store granularity sorting model based on DNN (Deep neural network) [4-6], with a limited number of store candidates (about 150) and the lack of more direct and important decision-making information such as commodities. Therefore, we regard commodity advertisement as the candidate supplement of stores, and open up the candidate space through the mixing of multiple commodities between stores, and the number of candidates can reach 1500+. In addition, considering the influence of advertising context and further expanding the scoring candidates to raise the upper limit of ranking, we upgraded the granularity of stores to the ranking of heterogeneous advertising combination granularity. Based on this, we built a generative advertising combination prediction system, and the candidate limit reached 1500X (considering the online performance, we finally chose 1500X). In the course of our exploration, we encountered three major challenges:

Product granularity estimation performance pressure: sinking down to the product granularity increases the number of candidates by at least 10 times, resulting in an increase in time that online estimation services cannot afford.
Difficult to model the relationship between combinations: It is difficult to describe the contextual relationship between stores and combined goods using Pointwise-Loss modeling.
Cold start problem of commodity advertisement: it is easy to form Matthew effect by using only the candidate exposed after model selection.

In view of the above challenges, the technical team carried out the following targeted optimization after thinking and practice:

High-performance heterogeneous mixed arrangement system: store information transfer and learning through BIAS network, so as to achieve high-performance commodity granularity prediction.
Generative advertising combination forecasting system: the product forecasting process is upgraded to list combination forecasting, and the context joint model is proposed to model the product context information.
Heterogeneous advertising cold start optimization: E&E (Exploit&Explore) optimization based on Thomson sampling algorithm to explore user interests in depth.

At present, high-performance heterogeneous and generative advertising combination has been estimated to be implemented in multiple advertising scenarios. Depending on the business of the scenario, it has increased by 4%-15% in the RPM (Revenue Per Mille) index, which measures advertising Revenue. Heterogeneous advertising cold start optimization takes effect in each business, and 10% randomness is given to the flow on the premise that the accuracy does not decrease. Our specific practices will be introduced in detail below.

2. Technical exploration and practice

2.1 High-performance heterogeneous hybrid system

After the grading granularity is lowered from stores to commodities, the number of candidates for ranking increases from 150 to 1500+, which brings about the improvement of ranking potential. Meanwhile, if the commodity prediction is conducted directly using the store model, it will bring an unaffordable increase in online time. Through analysis, we found that all goods under the store share the basic features of the store, occupying more than 80% of the network computing, but for multiple goods, only one calculation is required, and the unique and independent features of goods only occupy 20% of the network computing. Therefore, based on this feature, we refer to the practice of combinatorial estimation [7] to realize the heterogeneous mixed-platoon network. The high complexity store representation of the main network realizes the reuse of the output layer of the store network through the transfer learning of the common expression, so as to avoid the double calculation of the store network in the commodity estimation.

As shown in Figure 4 below, the whole network is divided into store network and commodity network. In the offline training stage, store network (main network) takes store characteristics as input, obtains store output layer, calculates store Loss, and updates store network; Commodity network (BIAS network) takes commodity characteristics as input to obtain commodity output layer, CONCAT operation is performed with store vector of output layer of store network, and then final commodity Loss is calculated, and store network and commodity network are updated at the same time.

In order to realize the reuse of store network output layer during online estimation, we feed the goods into the model in the form of List to achieve the scoring service once requested and obtain 1(store)+ N (commodity) estimated value. In addition, for the problem that the number of goods in stores is not fixed, we ensure dimension alignment by means of dynamic transformation of dimensions. Achieving a 10-fold increase in weight while maintaining the network size, while increasing the request time by only 1%.

Through heterogeneous mixed network, we get the estimated value of stores and each commodity under performance constraints. However, since advertising export is still based on stores as a unit for billing ranking, we need to apply the estimated value according to the characteristics of different business scenarios. For the convenience of description, “P store” is used below to represent the estimated value of stores, and “P commodity _I” represents the estimated value of the ith commodity.

Filter channel pages for competing heterogeneous ads

There are two types of display in the filter channel page, store and product, and the winning type of AD will finally be shown. In the training phase, each exposure is a sample, and a sample is one of the types of goods and stores. The store sample only updates the store network, while the merchandise sample updates both the store network and the merchandise network.
In the estimation stage, the click probability of stores and commodities is mutually exclusive, and we use Max operator: Max(P store,P commodity _1,… ,P commodity _n), if the store wins, the store information will be displayed, and the estimated value of the store will be used for downstream billing ranking; If any commodity wins, the commodity information is displayed, and the estimated value of the commodity is used downstream.

The combination relation of home page search is heterogeneous advertisement

Each display unit in the sorting list page of the home page search is composed of stores and two commodities, and the mechanism module conducts billing sorting for this display unit. During the training phase, each exposure is multiple samples: one store sample and multiple product samples. The store sample only updates the store network, while the merchandise sample updates both the store network and the merchandise network.
In the estimation stage, the Top2 products will be displayed by default before users click “more offers”, so they can choose the Top2 products with the highest estimated value as the display products, and the rest of the products will be sorted according to the estimated value. We need to estimate pCTR (goods store | goods | 1 2). From the Angle of mathematical analysis, we forecast stores or goods 1 or 2 by clicking the probability of goods, so we use the laws of probability addition operator: pCTR (goods store | goods | 1, 2) = 1 – (1 – P stores) * (1 – P commodity _1) * (1 – P commodity _2). Therefore, after obtaining the estimated value of stores and goods, the first thing is to sort the goods according to the estimated value to get the display order of goods, and select the Top2 estimated value of goods and store estimated value to calculate the probability addition rule, and get the estimated value of display units for the ordering and charging of stores.

Although the overall architecture of the system is similar, the sample generation methods are also different due to different usage scenarios, and the final output of P products of the model has different physical meanings. In competitive advertising, P products are used as another type of display juxtaposed with stores. In the combinatorial relationship advertisement, P product is the supplement of store advertisement display information, so it also has different application methods of estimated value. In the end, the high-performance heterogeneous hybrid system is implemented in multiple advertising scenarios. Depending on the business of the scenario, the RPM improvement range is between 2% and 15%.

2.2 Generative advertising mix estimation system

The click-through rate of an item in a list is affected not only by the quality of the item itself, but also by how it is displayed up and down. For example, when the context quality of an item is higher, the user is more likely to click on the context of the item, while when the context quality of the item is lower, the user is more likely to click on the context of the item, and this decision difference accumulates in the training data, thus forming a context bias. Eliminating context bias in training data is conducive to better positioning user intentions and maintaining the ecology of advertising system. Therefore, we refer to the idea of list ranking [8-9] to build a generative commodity ranking system and model commodity context information.

The context signal can be obtained by estimating the full permutation of the list of items, but the full permutation is very large (the full permutation of item candidate 10 is scored 10! = 21772800). In order to obtain the context signal with time permitting, we use quadratic estimation to prune the whole permutation result. In the first estimation, the Base model is used to score, and only Top N commodities are selected for ranking. In the second estimation, the context model is used to score all results of ranking. Turn the full array of dozen components from 10! To reduce the N! (On line, we chose N to be 3).

However, the secondary estimation will bring unbearable RPC time to the service. In order to go online under the constraint of performance, we implemented the secondary estimation module in TensorFlow. As shown in Figure 5 below, we finally achieved a high-performance combined prediction system based on pruning, with the overall time being equal to baseline.

By pruning and TF operators, any commodity input can sense its context signal. In order to model context information, we propose a context adaptive awareness model based on Transformer. The model structure is shown in Figure 6:

We first get store Emb and commodity Emb through Embedding layer respectively, and then get positional vector and estimation of positional vector through full link layer.
By combining the vector of non-ranked goods with the signal of ranked goods, Transformer models the contextual information of the goods and obtains the commodity Emb containing the contextual information.
The Emb of commodity containing context information and rank signal are spliced again, and the final output estimate of commodity containing context information and rank information is obtained through DNN nonlinear crossover. By strengthening the cross between goods, the purpose of modeling the context of goods was achieved, and the result of the generated advertising mix was estimated to be improved by RPM+2% in the homepage search.

2.3 Optimization of heterogeneous advertising cold start

In order to avoid Matthew effect, we will also take the initiative to test users’ new interest points and actively recommend new products to explore potential high-quality products. Before the model went online, we mined the products that users were interested in through random display. But an opportunity for you to show to users is limited, to show the user history like commodities, and explore new interest users will take up valuable opportunity, in addition, completely random display from CTR/PRS effect will be more obvious drop, so we consider a more reasonable way to solve the problem of “exploration and utilization”.

Compared with the traditional RANDOM display E&E algorithm, we adopt the Exploration algorithm based on Thompson sampling [10], which can reasonably control the accuracy loss and avoid the bias problem of bucket division for Exploration due to partial flow. Thompson sampling is a classical heuristic E&E algorithm. The core idea can be summarized as follows: give low randomness to goods with more Historical Impressions (HI, Historical Impressions), and give higher randomness to goods with less Historical Impressions. Specifically, we subject the pCTR to a beta(a,b) distribution, where:

${\frac{{a}}{{a+b}}= P}, a+b = n$

Where P is the function of pCTR and n is the function of EI. As a rule of thumb, the final function we use is:

${n = hyperN * \ text {} (log \ mathop {{}} \ nolimits_ {{10}} \ text {(+ 10 \} HI text {})) \ mathop {{}} \ nolimits ^ {{2}}}, * p = hyperP pCTR$

We control the randomness of the final results by adjusting hyperP and hyperN parameters. As shown in Figure 7 below, the mean value of action1 distribution is higher than that of Action2, and the randomness of ACTION3 distribution is stronger than that of the other two. High randomness may lead to a decrease in accuracy. We determined the full version of the hyperparameter through offline simulation of parameters. On the premise that the accuracy and effect of the final online model are not decreased, the goods displayed have 10% randomness.

2.4 Service Practices

Heterogeneous arrangement and advertising mix estimation effectively solve the problem of fewer store candidates under LBS restriction. For the two types of typical heterogeneous advertising introduced above: competitive heterogeneous advertising and combination heterogeneous advertising, we have implemented the corresponding technical exploration according to their display style and business characteristics, and achieved certain effects. As shown in Figure 8 below:

3 summary

This paper introduces the exploration and practice of heterogeneous advertising mix in meituan’s in-store search advertising business. We use high-performance heterogeneous mix network to meet performance challenges and apply heterogeneous prediction according to business characteristics. To modeling the advertisement context information, we will be the forecast process by single point forecast upgrade for the combination forecast model, and puts forward the context combination forecast model, the modeling of goods places and context information, and then, by E&E strategy based on Thompson algorithm for goods cold start problem is optimized, has obtained certain achievements in multiple scenarios. Recently, more and more business scenarios have begun to upgrade the display style, for example, the food category is adjusted from stores to dishes advertising, the hotel category is adjusted from stores to room display. The schemes and technologies mentioned in this paper are also in the process of gradual promotion and implementation.

It is worth mentioning that, compared with Meituan which takes stores as the main body of advertising, the main body of advertising in the industry is commodities and content. The skills of common expression migration and generative combination prediction mentioned in this paper can be applied to the combination of commodities and creativity to further expand the scale of candidates.

The heterogeneous and mixed advertising project is also an important attempt to break the original iterative framework from a business perspective. We want the project to solve business problems through technical means, and then push back technological advances through business understanding. In addition, we will also conduct more exploration on advertising candidates, looking for new breakthrough points, so as to further design a more perfect network structure, and constantly release the potential space of the ranking system.

4 Reference Materials

[1] Huang, Jianqiang, et al. “Deep Position-wise Interaction Network for CTR Prediction.” Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2021.
[2] Qi, Yi, et al. “Trilateral Spatiotemporal Attention Network for User Behavior Modeling in Location-based Search.” Proceedings of the 30th ACM International Conference on Information & Knowledge Management. 2021.
[3] Hu Ke, Strong. Breakthrough and imagination of advertising depth estimation technology in meituan store scene
[4] Cheng, Heng-Tze, et al. “Wide & deep learning for recommender systems.” Proceedings of the 1st workshop on deep learning for recommender systems. 2016.
[5] Zhou, Guorui, et al. “Deep interest network for click-through rate prediction.” Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining. 2018.
[6] Ma, Jiaqi, et al. “Modeling task relationships in multi-task learning with multi-gate mixture-of-experts.” Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 2018.
[7] Gong, Yu, et al. “Exact-k recommendation via maximal clique optimization.” Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining. 2019.
[8] Guo, Huifeng, et al. “PAL: a position-bias aware learning framework for CTR prediction in live recommender systems.” Proceedings of the 13th ACM Conference on Recommender Systems. 2019.
[9] Feng, Yufei, Roche Recommender System in the Permutation Prospective.” arXiv Preprint arXiv:2102.12057 (2021).
[10] Ikonomovska, Elena, Sina Jafarpour, and Ali Dasdan. “Real-time bid prediction using thompson sampling-based expert selection.” Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. 2015.

Recruitment information

The algorithm team of Meituan in-store business group advertising platform is fully responsible for the optimization of advertising algorithms for all in-store businesses, and continues to improve the cash efficiency of commercial traffic on the premise of ensuring user experience and ROI of advertising merchants. The main technical directions include trigger strategy, quality estimation, mechanism design, creative generation, creative optimization, anti-cheating, business strategy, etc. The strong technical atmosphere of the team drives the sustainable development of the business through continuous breakthroughs in cutting-edge technologies. The team focuses on talent cultivation and has a perfect and mature training mechanism to help everyone grow up quickly.

Post requirements

At least 2 years relevant working experience, familiar with common machine learning principles and deep learning models, have CTR/CVR/NLP/CV/RL model practical experience.
Have excellent analytical and problem-solving skills, keep learning ability and curiosity about new things, and be passionate about solving challenging problems.
Good programming skills, solid foundation of data structure and algorithm, familiar with Python/Java/Scala/C++ or above.
Bachelor degree or above in computer science, automation, electronic information, mathematics or related field.

The following requirements are preferred

Internet advertising/search/recommendation related work experience.

If you are interested, please send your resume to [email protected] (please mark the email title: Guangping Algorithm Team).

Read more technical articles from meituan’s technical team

| in the public bar menu dialog reply goodies for [2021], [2020] special purchases, goodies for [2019], [2018] special purchases, 【 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 mark “Content reprinted from Meituan Technical team”. This article shall not be reproduced or used commercially without permission. For any commercial activity, please send an email to [email protected] for authorization.