Exploration and Practice of Heterogeneous Advertising Mixed Ranking in Meituan's Store-to-Store Business

Under the distance constraint of LBS (Location Based Services), fewer candidates restrict the potential space of the entire in-store advertisement ranking system. This paper introduces our candidate extension from the perspective of candidate types, and uses a high-performance heterogeneous shuffling network to meet the performance challenge, thereby improving the potential upper limit of the local life scene ranking system. I hope it can inspire students who are engaged in related fields.

1 Background and Introduction

1.1 Background

Meituan's in-store advertising is responsible for the commercial realization of Meituan's search traffic, and serves many local life service merchants such as in-store catering, leisure and parent-child entertainment, beauty medical beauty, hotel tourism, etc. The quality estimation team is responsible for the quality score estimation of / ^{and customer unit price/transaction volume in the advertising system 2] and other innovative technologies have made some breakthroughs CTR estimation problems in terms of users and .}

However, the above papers focus on model accuracy, and model accuracy and advertisement candidates jointly determine the quality of the ranking system. However, from the perspective of advertising candidates, compared with the candidate sets of traditional e-commerce, search advertising is limited by LBS (Location Based Services, location-based services), so there are fewer candidates for stores in some categories, while candidates Less and seriously restrict the potential space of the entire sorting system . When the traditional method of increasing the number of candidates fails to obtain benefits, we consider expanding and optimizing the advertising candidates in order to increase the potential upper limit of the local life scene ranking system.

1.2 Scenario introduction

A single store advertisement is not enough to satisfy the fine-grained intention of users to find products and services. In some scenarios, commodity advertisements are used as candidate supplements for store advertisements, and the two determine the display advertising style in a competitive manner; in addition, there are also some scenarios where commodity advertisements are displayed in combination with store advertisements in the form of hanging down. Various forms of heterogeneous advertising display styles have brought opportunities and challenges to the in-store advertising technology team. According to the characteristics of the business scenario, we have optimized the heterogeneous advertising in a targeted manner. The following takes the Meituan marriage channel page and the Meituan homepage search as examples to introduce two types of typical heterogeneous mixed ads: competitive heterogeneous ads and combination heterogeneous ads.

• Heterogeneous advertisements with competitive relationship: two types of advertisements of stores and products are competing and shuffled, and the advertising display type is determined by comparing the pCTR in the shuffling model. As shown in Figure 1 below, the first place in the left column is the winner of the store type advertisement, and the displayed content is the store image, store title and the number of star reviews of the store; the first place in the right column is the winner of the product type advertisement, and the displayed content is the product image, product title and corresponding store. . The advertising system determines the arrangement order and display type of advertisements, and when the commodity type advertisement wins, the system determines the specific commodity to be displayed.

• Combination Relationship Heterogeneous Advertising: The store advertisement and its product advertisement are combined into a display unit (blue box) for list sorting, the product belongs to the store, and the two types of heterogeneous advertisements are combined to display . As shown in Figure 2 below, the store advertisement displays the store's header image, title price and other information; two product advertisements display the product price, title and sales information. The advertising system determines the arrangement order of the display units, and determines the Top 2 products displayed in the store's product collection.

1.3 Introduction to Challenges and Practices

At present, the online search advertising model is a store granularity ranking model based on DNN (Deep Neural Network) ^[4-6] . The number of store candidates is limited (about 150) and more direct and important decision-making information such as products is missing. Therefore, we use product advertisements as the candidate supplement of the store, and open the candidate space through the mixed arrangement of multiple products under the store and the store, and the number of candidates can reach 1500+. In addition, considering the influence of advertising context and further expanding the scoring candidates to increase the upper limit of sorting, we upgraded the granularity of stores to the granularity of heterogeneous advertising portfolios, and built a generative advertising portfolio estimation system based on this. The candidate limit reached 1500X (considering line On the performance we finally choose 1500X). In the process of exploration, we encountered the following three challenges:

• Commodity granularity prediction performance pressure : After sinking to commodity granularity, the number of candidates will increase by at least 10 times, resulting in an unbearable increase in time consumption for online prediction services.
• Difficulty modeling the relationship between combinations : The context relationship between the store and the combined product is difficult to describe using Pointwise-Loss modeling.
• Commodity advertising cold start problem : Only the candidates exposed after model selection are used, which is easy to form the Matthew effect.

In response to the above challenges, the technical team has made the following targeted optimizations after thinking and practice:

• High-performance heterogeneous shuffling system : Transfer learning of store information through bias network to achieve high-performance commodity granularity prediction.
• Generative advertising combination prediction system : Upgrade the commodity prediction process to list combination prediction, and propose a context joint model to model commodity context information.
• Heterogeneous advertising cold start optimization : E&E (Exploit&Explore, exploration and utilization) optimization based on Thomson sampling algorithm to deeply explore the interests of users.

At present, high-performance heterogeneous mix and generative advertising mix estimates have been implemented in multiple advertising scenarios. Depending on the scenario business, the RPM (Revenue Per Mille) indicator that measures advertising revenue has improved. 4%~15%. Heterogeneous advertising cold-start optimization takes effect in each business, giving 10% randomness to traffic without reducing accuracy. Our specific approach will be described in detail below.

2 Technology exploration and practice

2.1 High-performance heterogeneous shuffling system

After the scoring granularity is lowered from the store to the product, the number of candidates for sorting increases from 150 to 1500+, which increases the sorting potential. At the same time, if the store model is used to directly estimate the product, it will bring unaffordable time to online. Increase. Through analysis, we found that all commodities in the store share the basic characteristics of the store, occupying more than 80% of the network calculation, but only need to calculate once for multiple commodities, while the commodity characteristics unique to the commodity and need to be calculated independently only occupy 20% of the network calculation. network computing. Therefore, based on this feature, we refer to the practice of combining and predicting ^[7] to implement a heterogeneous shuffled network. The high-complexity store representation of the main network realizes the reuse of the output layer of the store network through transfer learning of common expressions, thereby avoiding the repeated calculation of the store network during product estimation.

As shown in Figure 4 below, the entire network is divided into a store network and a commodity network. In the offline training stage, the store network (main network) takes the store features as input to obtain the output layer of the store, calculates the store Loss, and updates the store network; the product network (bias network) takes the product features as the input, obtains the product output layer, and communicates with the store. The output layer store vector of the network performs the CONCAT operation, then calculates the final product Loss, and updates the store network and the product network at the same time.

In order to realize the multiplexing of the output layer of the store network during online estimation, we feed the products into the model in the form of List, realize the request for a scoring service, and obtain 1 (store) + n (product) estimated values. In addition, for the problem that the number of products in the store is not fixed, we ensure the alignment of dimensions through dynamic dimension transformation. To achieve a 10-fold increase in the weight of the network while maintaining the network scale, the request time is only increased by 1%.

Through the heterogeneous mixed network, we obtained the estimated value of the store and each product under the performance constraints, but since the advertising export still uses the store as the unit for billing and sorting, we need to apply the estimated value according to the characteristics of different business scenarios. . For the convenience of description, hereinafter, "P store" is used to represent the estimated value of the store, and "P product_i" is used to represent the estimated value of the ith product.

Filtering channel pages for competition with heterogeneous ads

• In the filter channel page, there are two display types of store and product to compete, and the winning advertisement type will be displayed finally. In the training phase of , each exposure is a sample, and a sample is one of the types of products and stores . The store sample only updates the store network, and the product sample updates both the store network and the product network.
• In the estimation stage, the click probability of the store and the product are mutually exclusive. We use the Max operator: through Max(P store, P product_1 ,...,P product_n ), if the store wins, the store information will be displayed, and the store will be displayed. The estimated value of the product is used for downstream billing sorting; if any product wins, the product information will be displayed, and the estimated value of the product will be used downstream.

Home Search Combination Relationship Heterogeneous Advertising

• Each display unit in the sorting list page of the homepage search consists of a store and two commodities, and the mechanism module performs billing sorting for this display unit. In the training phase of , each exposure consists of multiple samples: one store sample and multiple commodity samples . The store sample only updates the store network, and the product sample updates both the store network and the product network.
• In the estimation stage, since the Top 2 products are displayed by default before the user clicks [More Offers], the Top 2 products with the highest estimated value can be selected as the displayed products, and the rest of the products are sorted by the estimated value. We need to estimate pCTR(Store|Item1|Item2). From a mathematical point of view, we are estimating the probability that a store or item 1 or item 2 is clicked, so we use the probability addition rule operator: pCTR(Store|Item1|Item2) = 1 - (1-PStore) (1-P Commodity_1) (1-P Commodity_2). Therefore, after obtaining the estimated value of the store and the product, first sort the products according to the estimated value, get the display order of the products, and select the estimated value of the top 2 products and the estimated value of the store to calculate the probability addition method to get the display. The estimated value of the unit is used for store sorting billing.

Although the overall architecture of the system is similar, due to different usage scenarios and different sample generation methods, the P products finally output by the model have different physical meanings. In the competitive relationship advertisement, the P product is another display type that is side by side with the store; in the combined relationship ad, the P product is a supplement to the store advertisement display information, so there are different ways of applying the estimated value. In the end, the high-performance heterogeneous shuffling 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

In the product list, the click-through rate of the product is not only affected by its own quality, but also affected by the products displayed above and below it. For example, when the context quality of the item is higher, users are more inclined to click the context of the item, and when the context quality of the item is lower, the user is more inclined to click the item, this decision difference will accumulate in the training data, forming contextual bias. Eliminating the contextual bias in the training data is conducive to better locating user intentions and maintaining the ecology of the advertising system. Therefore, we refer to the idea of list sorting ^[8-9] to build a generative commodity sorting system and model Product context information.

The context signal can be obtained by estimating the full permutation of the commodity list, but the full permutation score is extremely large (the full permutation score of 10 commodity candidates is 10!=21,772,800). In order to obtain the context signal when time-consuming allows, we prune the full permutation results by means of quadratic estimation. In the first estimation, the Base model is used to score, and only Top N products are ranked. In the second estimation, the context model is used to score all the results of the arrangement. Reduce the total permutation hit size from 10! to N! (online, we choose N to be 3).

However, the secondary estimation will bring unbearable RPC time to the service. In order to go online under the constraints of performance, we have implemented the secondary estimation module in TensorFlow. As shown in Figure 5 below, we finally realized a high-performance combined prediction system based on pruning, and the overall time consumption was the same as the baseline.

Through pruning and TF operator, any commodity input can perceive its context signal. To model contextual information, we propose a Transformer-based context-adaptive awareness model. The model structure is shown in Figure 6:

1. We first obtain the store Emb and the product Emb by passing the store features and product features through the Embedding layer respectively, and then obtain the non-ranked product vector and the non-ranked estimated value through the full link layer;
2. Splicing the non-rank commodity vector with the commodity rank signal, modeling the context information of the commodity through the Transformer, and obtaining the commodity Emb containing the context information;
3. The commodity Emb containing the context information and the rank signal are spliced again, and the final output commodity estimated value containing the context information and the rank information is obtained through the non-linear intersection of DNN. By strengthening the intersection between products and achieving the purpose of modeling the product context, the final generative advertising combination is estimated to achieve an RPM+2% increase in the homepage search effect.

2.3 Heterogeneous advertising cold start optimization

In order to avoid the Matthew effect, we will also take the initiative to test users' new points of interest, and actively recommend new products to discover potential high-quality products. Before the model goes live, we mine products that users are interested in by randomly displaying them. However, the opportunity to display to users is limited. Displaying products that users like in history and exploring new interests of users will take up valuable display opportunities. In addition, completely random display will have a relatively obvious decline in terms of CTR/PRS and other effects. So we consider solving the problem of "exploration and utilization" in a more reasonable way.

Compared with the traditional random display E&E algorithm, we use the Exploration algorithm ^[10] based on Thompson sampling, which can reasonably control the loss of accuracy and avoid the bias problem of Exploration bucketing due to partial traffic. Thompson sampling is a classic heuristic E&E algorithm. The core idea can be summarized as follows: lower randomness is given to commodities with more historical exposures (HI, Historical Impressions), and higher randomness is given to commodities with less historical exposure. Specifically, we make the estimated value of the commodity (pCTR) obey a beta(a,b) distribution, where:

$${\frac{{a}}{{a+b}}=p} ，a + b = n $$

where p is a function with pCTR as an independent variable, and n is a function with EI as an independent variable. As a rule of thumb, the function we ended up using was:

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

We control the randomness of the final presentation results by adjusting the hyperP and hyperN parameters. As shown in Figure 7 below, action1 has a higher mean distribution than action2, and action3 is more random than the other two distributions. Higher randomness may lead to a decrease in accuracy. We determine the hyperparameters of the full version through offline simulation of the parameters. Under the premise that the accuracy and effect of the final online model does not decrease, the displayed products have 10% randomness.

2.4 Business Practice

Heterogeneous shuffling and advertising mix estimation effectively solve the problem of fewer store candidates under LBS constraints. For the two types of typical heterogeneous advertisements introduced above: heterogeneous advertisements with competitive relationship and heterogeneous advertising with combined relationship, we have implemented corresponding technical explorations according to their display styles and business characteristics, and achieved certain results. As shown in Figure 8 below:

3 Summary

This paper introduces the exploration and practice of heterogeneous ad mix in Meituan's in-store search advertising business. We use a high-performance heterogeneous mix network to meet performance challenges, and apply heterogeneous estimates according to business characteristics. In order to model the contextual information of advertisements, we upgrade the product prediction process from single-point prediction to combined prediction mode, and propose a contextual combined prediction model to model the product ranking and contextual information. Then, through the Thompson-based algorithm The E&E strategy optimizes the cold start problem of commodities, and has achieved certain results in multiple scenarios. Recently, more and more business scenarios have begun to upgrade the display style. For example, the food category has been adjusted from stores to dish advertisements, and the hotel category has been adjusted from stores to room display. The solutions and technologies mentioned in this article are also gradually being promoted. in the process of landing.

It is worth mentioning that, compared with Meituan, which uses stores as the main body of advertising, the main body of advertising in the industry is mainly products and content. The skills of expression migration and generative combination estimation mentioned in this article can be applied to products and creatives. On the combination problem of , further expand the candidate size.

The project of heterogeneous mixed arrangement of advertisements is also an important attempt to break the original iterative framework from a business perspective. We hope that this project can solve business problems through technical means, and then reverse the progress of technology through business understanding. In addition, we will also conduct more explorations on the advertising candidate problem to find new breakthrough points, so as to further design a more complete network structure and continuously release the potential space of the ranking system.

4 References

• [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, Jianqiang, etc. The breakthrough and imagination of advertising depth estimation technology in the scene of Meituan to the store
• [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, et al. "Revisit 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.

Job Offers

The advertising platform algorithm team of Meituan's in-store business group is fully responsible for the optimization of advertising algorithms for all in-store-related businesses. On the premise of ensuring user experience and ROI of advertising merchants, it will continue to improve the monetization efficiency of commercial traffic. The main technical directions include trigger strategy, quality estimation, mechanism design, creative generation, creative optimization, anti-cheating, business strategy, etc. The team has a strong technical atmosphere, and drives continuous business development through continuous breakthroughs in cutting-edge technologies. The team regards talent cultivation and has a complete and mature training mechanism to help everyone grow rapidly.

job requirements

• More than two years of relevant work experience, familiar with common machine learning principles and deep learning models, and have practical experience in CTR/CVR/NLP/CV/RL models.
• Possess excellent analytical and problem-solving skills, maintain the ability to learn new things and be curious, and be passionate about solving challenging problems.
• Possess good programming ability, solid data structure and algorithm foundation, familiar with Python/Java/Scala/C++ two or more languages.
• Bachelor degree or above in computer, automation, electronic information, mathematics or related majors.

The following conditions are preferred

• Internet advertising/search/recommendation related work experience in a certain field.

Interested students can submit their resumes to: [e-mail] (please specify the subject of the email: Guangping Algorithm Team).

Read a collection of more technical articles from the Meituan technical team

the front | algorithm | backend | data | security | operation and maintenance | iOS | Android | test

| reply to keywords such as [2021 stock], [2020 stock], [2019 stock], [2018 stock], [2017 stock] in the public account menu bar dialog box, you can view the collection of technical articles by the Meituan technical team over the years.

| This article is produced by Meituan technical team, and the copyright belongs to Meituan. Welcome to reprint or use the content of this article for non-commercial purposes such as sharing and communication, please indicate "The content is reproduced from the Meituan technical team". This article may not be reproduced or used commercially without permission. For any commercial activities, please send an email to tech@ meituan.com to apply for authorization.