Advertising and marketing of Meituan in the O2O scenario

As the largest online local life service platform in China, Meituan covers catering, hotels, travel, leisure and entertainment, takeout delivery and other life scenes, connecting hundreds of millions of users and millions of merchants. One of meituan’s core issues is how to help local merchants carry out online marketing so that they can quickly and effectively reach target user groups and improve operation efficiency. Machine learning related technologies play a very key role in local online marketing scenarios.

This paper will introduce it from five aspects. Firstly, it introduces the characteristics of advertising business in O2O scenario, and the differences with B2B and B2C advertising business. Secondly, it introduces the most important indicators of O2O advertising business from three dimensions of merchant effect perception, user experience and media platform revenue. Thirdly, based on the business characteristics and indicators described in the previous two sections, the mechanism design of online advertising marketing in the O2O scenario is introduced. Fourthly, it introduces the push advertisement under the O2O special real-time scenario; Finally, a brief introduction of O2O advertising system related tools.

Advertising business features in O2O scenarios

Before the rise of O2O business model and related platforms, businesses of big brands rely on the following ways to promote brand awareness due to their large income scale and sufficient marketing expenses: traditional media (such as TV, radio and newspaper); Internet traffic (such as traditional search engines, portal websites, etc.); Outdoor advertising (such as bus and subway body, billboard, light box, etc.). Through the above media, merchants can quickly reach a large number of users and promote the image of the brand. The marketing approach has its limits. First of all, the capital threshold of advertising is high, and merchants with limited marketing budgets cannot afford the relevant expenses; Secondly, for direct effect-oriented businesses, the above delivery forms are too extensive and cannot form a direct closed-loop purchase conversion effect. For the majority of small and medium-sized businesses, their marketing budget is limited and they pay more attention to direct purchase transformation, the main way to obtain potential customers is to distribute leaflets, distribute gifts, along the street loudspeaker advertising. However, the coverage of these offline marketing methods to potential consumers is relatively limited, and these methods cannot be sustained for a long time.

O2O local life service platforms represented by Meituan have grown rapidly and gradually become one of the most important means of online marketing for local service merchants. Meituan is home to hundreds of millions of consumers who use the platform to find merchants, check out deals and browse reviews. For merchants, they are the most direct potential consumers. Through online marketing on Meituan platform, merchants can gain more display opportunities to attract more customers to their stores. With the help of convenient online consulting, booking and payment means, the advertising business on the platform can form a closed loop of effect, so that merchants can clearly and accurately grasp the effect of advertising and optimize the advertising strategy.

For Meituan, based on the mining and analysis of users’ big data, the platform can connect users’ online and offline behaviors, understand and judge users’ emotions, attitudes and needs, and provide users with real-time, directional and creative information and content services under specific scenarios composed of time, place, users and relationships. Compared with traditional B2C and B2B business models, online marketing advertising in the O2O scenario has its unique attributes, which are mainly reflected in four dimensions: mobile, localization, scenarioization and diversity.

With the rapid development of broadband wireless access technology and mobile terminal technology, people gradually began to use mobile phones and other mobile devices to obtain information and services from the Internet anytime and anywhere. In this era, whether it’s reading news, social communication or e-shopping, people are used to meeting their needs directly through mobile apps. In fact, Meituan took the initiative to adapt to the historical trend in the early stage of the development of mobile Internet and vigorously developed mobile service capabilities. At present, more than 90% of transactions are made through mobile Internet services. As a localized life service marketing model connecting people and services, O2O advertising has distinct characteristics of mobile and localization.

• Mobile. It is mainly embodied in three aspects: accuracy, immediacy and interactivity. With sensors on mobile devices, we can accurately understand users’ geographical location and push more accurate ads. Most users carry their phones with them at all times, so advertising messages can be pushed to users in a timely manner. A variety of powerful mobile applications provide a variety of interactive possibilities for advertising. For example, on Meituan App, users can directly complete the information inquiry, queuing and transaction of merchants.

• Localization. Transformation effect-oriented O2O advertising marketing, the target users of marketing are the people near the local businesses that provide services. On Taobao, a pair of leather shoes can be promoted and sold to users all over the country. No matter where consumers are, logistics and express delivery will accurately deliver the goods to the hands of consumers. On Meituan, a hotpot restaurant in Wudaokou is best marketed to diners in the vicinity of wudaokou, who are most likely to come directly to the restaurant. In fact, by looking at actual transaction data, we found that in over 90% of transactions, the distance between the user and the merchant was less than 3 km. In order to achieve good results, marketing activities must be targeted to select target groups. In O2O advertising, target groups are localized user groups. The precise positioning of mobile devices ensures that merchants can find their target audience.

• The scene. Consumers, mobile devices, time and space constitute the precise scene of user consumption demand. In the PC era, users’ identifiers take cookies as the carrier, but cookies are easy to clear. Meanwhile, a computer may be used by many people, which makes it difficult to effectively connect user information, and even basic information such as the age and residence of the audience cannot be accurately grasped. However, in the era of mobile Internet, we can analyze and reconstruct the attributes and preferences of users in all aspects by analyzing and mining the various behavioral footprints left by users on the platform, so as to produce very accurate user portraits. Under the premise of understanding user information such as geographic location, consumption intention and behavior trajectory, O2O advertising marketing can provide real-time, targeted and creative marketing content for users under specific scenarios composed of time, place, users and demands, and connect users’ online and offline behaviors. For example, on a sunny afternoon, for a white collar who works in CBD and has the habit of drinking afternoon tea, the platform can timely push afternoon tea or coffee shop merchants.

• Diversity. O2O business model is faced with a variety of local life service businesses, different businesses have different characteristics, and put forward different needs for O2O advertising and marketing. To take a simple example, different service businesses have different local requirements for target users: catering services are sensitive to distance, and the target user groups of these service merchants are diners around the merchants; Less sensitive to distance are services such as wedding photography, which target newlyweds across the city.

The interests of merchants, users and platforms are balanced

Advertising system and search system, recommendation system, has a very similar system architecture: they mostly use the search and sorting process system. Based on this, there are many people who believe that the advertising business is no different from the search recommendation business. In fact, the advertising business has its own unique laws. Advertising is first and foremost a commercial activity that predates the Internet. As a commercial activity, the interests of merchants, consumers and media platforms should be taken into account. These interest indicators are the guiding light for the sustainable and healthy development of advertising business. This section will analyze the effect perception of merchants, user experience and platform revenue in MEituan O2O advertising marketing from the perspective of business activities.

Merchant effect perception

The fundamental purpose for merchants to conduct advertising marketing on Meituan advertising platform is to reach more potential consumers through Meituan and obtain the maximum incremental benefits. The cost of a local life service type business can be divided into two parts: variable cost and fixed cost. Variable cost is the cost that changes linearly with the volume of business, mainly from the consumption of raw materials. The fixed cost is the cost that will not change with the change of business volume in a certain period, such as the input of facade decoration, the rent of the store, the basic salary of the store service personnel, etc. If merchants do not have enough business volume and cannot attract enough consumers, the cost of unit business volume will remain high, leading to serious losses. Therefore, for the catering industry, the primary goal of the business is to improve the table rate and reduce the vacancy rate, and for the hotel industry, the primary goal of the business is to improve the full room rate and reduce the vacancy situation. The existence of fixed cost is the basic premise for local merchants to carry out O2O advertising marketing.

From the perspective of merchants, the effect of O2O advertising marketing can be measured from three dimensions: advertising visibility, online incremental revenue brought by advertising and overall incremental revenue brought by advertising.

For merchants, visibility is the most preliminary and direct marketing result and the fastest effect feedback they can get. The visibility of advertisements indicates that merchants’ marketing information has begun to reach potential consumer groups through media platforms. Therefore, stable and reliable advertising display expectation is the most basic requirement to win the majority of merchants’ trust in O2O advertising marketing.

The online incremental revenue from advertising refers to the online revenue generated by advertising on meituan and other media platforms. This part of income can be divided into two categories: one is the income brought by direct online orders, such as group purchase, hotel reservation, etc.; The other is revenue from indirect transactions such as online bookings. For this part of revenue, the platform can provide accurate statistics, analysis and feedback to advertising merchants. For takeout, wedding photography, hotel tourism and other industries that are highly dependent on online traffic and online transactions, their online revenue accounts for a large proportion of the overall revenue, which directly reflects the situation of business activities of merchants.

In addition to direct online transactions, another scenario for users to use Meituan is to check the information of dishes, reviews and geographical locations of merchants through the platform, and then directly to the store to make purchases. The overall incremental revenue brought by advertising to merchants includes the revenue brought by offline customer drainage. Online transactions of catering businesses only account for a small part of the overall revenue of stores. Therefore, the measurement of advertising effectiveness needs to consider both online and offline revenue. Compared with online transaction revenue, it is difficult to make accurate statistics of part of revenue from offline traffic diversion. However, the platform can make accurate statistics of part of users’ in-store status through real-time geographical location of users, or estimate in-store data through exposure and click-to-in-store data funnel model. With the popularity of electronic payment in the future, the platform will be able to better calculate the overall income of merchants.

After understanding the main measurement index of O2O advertising marketing effect, Return over Investment (ROI), a commonly used evaluation index, is needed to determine whether the advertising cost of merchants is really low, that is, the proportion of the total output and total Investment of an advertising campaign. Corresponding to the two metrics of advertising revenue, ROI can also be divided into online payment ROI and overall ROI: Online payment ROI is equal to the online incremental transaction volume divided by advertising cost consumption, and overall payment ROI is equal to the overall store revenue increment divided by advertising cost consumption. With limited advertising budgets, merchants are always looking to optimize advertising and improve ROI.

The user experience

Effectively ensuring user experience is the basic prerequisite for Meituan to carry out O2O advertising marketing. The value of a platform can only be realized if it can ensure the user experience and be useful to users. By keeping more users and being active on the platform, Meituan can attract more local lifestyle service merchants for advertising and generate more traffic for advertising realization.

Meituan mainly designs and measures user experience indicators from short-term and long-term dimensions. Starting from the funnel of user behavior — information exposure, user clicks and user transactions, the short-term user experience index mainly considers clicks and transactions. The first short-term user experience metric is Click through Rate (CTR), which is mathematically expressed as the number of clicks divided by the number of impressions. CTR reflects the quality and relevance of merchant information displayed to users. Advertising information display unrelated to user intentions and not matched with the time and place scene of users can not meet user needs and attract users’ clicks, resulting in a low CTR. Click through rate is subdivided into the click through rate exposed to the advertisement and the click through rate of the whole page. The former measures the quality of the advertisement itself, and the latter reflects the influence of the advertisement on the overall information presentation effect (natural results plus advertising results). In addition to its own low click-through rate, inferior advertising will also disturb the overall browsing behavior of users, making users unable to obtain the local life service information they need joyfully.

In order to obtain real exposure, the proportion and time of each POI actually displayed on the mobile phone screen will be monitored at the buried point on the mobile terminal, and POI exceeding a certain display proportion and time threshold will be included in the statistics of exposure times.

The second short-term user experience metric is the Conversion Rate (CVR), which is mathematically expressed as the number of transactions (Order) divided by the number of clicks (Click). Conversion rate also reflects the relevance and quality of merchant information display. Merchant display that does not match user needs will fail to facilitate the completion of transactions, resulting in a lower conversion rate. Similar to CTR, CTR can be divided into AD CTR and overall page CTR. The advertising conversion rate is also directly proportional to the ROI of online transactions of merchants. Accurate and effective advertising can not only improve user experience, but also improve the ROI of merchants.

Long Term User Experience metrics measure the lasting impact of ads on users over a longer time horizon. Long-term user experience indicators mainly include two indicators: return rate and repurchase rate. Return rate is an indicator of long-term retention of users, which means whether users will log in and use Meituan platform again in a certain period of time. The return rate index includes weekly return rate, monthly return rate and so on. Low-quality advertising disturbs the users’ experience of convenient access to merchant information through the platform, which leads to the loss of users from the platform, thus reducing the return rate. The repurchase rate reflects the index of user consumption experience, and its significance is whether users will repurchase a merchant’s service in a certain period of time. Similarly, low-quality merchant services will damage users’ consumption experience and make users stop making the same consumption, thus leading to the decline of re-purchase rate.

In order to accurately measure the advertising effect of user experience, in addition to policy changes comparison test, the platform will retain a small portion of the flow rate as a control group for a long time, no part of this user to carry out the advertising, by comparing the overall flow and the differences in the relevant user experience indicators in the control group, to determine the long-term effects of advertising on the user experience. Then supervise and guide the platform to optimize the advertising strategy.

Platform earnings

As a media platform, The goal of Meituan is to optimize traffic realization efficiency and achieve the best connection between merchants’ marketing demands and users’ consumption demands while ensuring merchants’ ROI and user experience.

The previous two sections have introduced the basic concepts of merchant ROI and user experience. We know that only when ROI of merchants is guaranteed, more merchants and more budgets will be put into the advertising system. Only by ensuring user experience, can there be more users and more traffic for advertising realization. These two things determine the size of the cheese in the advertising business.

Traffic realization efficiency measures advertising revenue per unit of traffic. For display advertising business, traffic realization efficiency is mainly expressed by Revenue per Mille (RPM). For Search advertising, traffic realization efficiency is mainly expressed by Revenue per Search (RPS).

From the perspective of traffic supply, advertising Revenue is the product of advertisement exposure times, click rate and click price (CPC). From the perspective of traffic demand, advertising Revenue is the product of the number of advertisers and Average Revenue Per User (ARPU). On the premise that the number of advertiser accounts, budget and traffic are stable, the improvement of traffic realization efficiency is mainly driven by two key indicators, click rate and click unit price, and the benign improvement of these two indicators depends on the mechanism design and advertising algorithm, the details will be described below.

O2O advertising mechanism design

The characteristics of MEituan O2O advertising and marketing are introduced and the interests of merchants, users and platforms are analyzed. Based on the above characteristics and interests, this section will explain the design principle of O2O advertising mechanism in meituan’s actual business, including advertising space setting, advertising recall mechanism and advertising ordering mechanism.

Advertising space setting

On the mobile terminal, the natural results of Meituan are presented in the form of a list of information, while the advertisement occupies a fixed position in the list (range floating fixed position) for display. From the perspective of effect perception of merchants, fixed spot advertisement form can give merchants a relatively certain advertising display expectation, so that merchants have a clear bidding target (i.e. fixed display location).

The setting of advertising space needs to comprehensively consider and balance the interests of merchants, users and platforms. Too intensive advertising location design and advertising display will reduce the efficiency of users looking for merchant information and affect the user experience. Too sparse advertising location design leads to too few opportunities for advertising display, resulting in insufficient platform flow realization efficiency. Headspace has a big impact on user experience, but it can get more exposure, be more valuable, and motivate merchants to bid more. The advertising space at the waist and tail has little influence on user experience, but the probability of exposure of advertising space is small and it cannot effectively stimulate the bidding of merchants. On the one hand, the actual advertising space setting of Meituan takes into account the characteristics of each booth and business. On the other hand, A/B test is conducted to compare and select various schemes, and finally the design scheme that can effectively take account of user experience, merchant effect and platform revenue is selected.

Advertising recall mechanism

AD recalls are technically very similar to search and recommendations. The search scene advertisement will use the user’s query words to search for the matching merchant in the AD merchant index, and the recommendation scene advertisement will match the appropriate merchant according to the user’s intention, location and other scene information.

Query rewriting is an important technology in search advertisement matching. On the one hand, we use traditional natural language processing methods to effectively analyze the query (such as component analysis) and complete synonym and synonym rewriting. On the other hand, we use the deep semantic similarity neural network model (DSSM) and Sequence to Sequence model to rewrite queries to further improve the coverage and accuracy of advertising matching.

According to the characteristics of O2O business model and the interests of all parties in advertising business, the advertising recall mechanism has been optimized and improved based on the traditional search recommendation recall mechanism. We introduced the concept of layer-by-layer recall in recall, and set Match Level from tight to loose to control the quality of recall ads in each layer. In the case that sufficient number of advertising candidates have been recalled at the current relevance Level, no subsequent recall will be carried out. Correlation leveling considers a variety of correlation factors: query matching pattern, distance, and star rating. For example, for Query matching pattern, Query exact matching pattern will be used first in advertising recall, followed by fuzzy matching pattern, and semantic matching pattern will be used last. Considering the distance factor, the advertisement recall will give priority to the merchants within 3 kilometers, then select the merchants within 5 kilometers, and finally try to recall the whole city.

The setting of relevance level should fully consider the characteristics of different O2O businesses. For example, distance setting, for catering traffic, the system will give priority to recall merchants within 3 kilometers, and for distance relatively insensitive wedding photography traffic, the system will relax restrictions, give priority to recall merchants within 10 kilometers, or directly adopt a city-wide recall strategy.

Advertising ranking mechanism

Like traditional search advertising, meituan’s ads are Cost Per Click (CPC) ads, in which advertisers bid based on the value of an AD’s clicks, and the AD system ranks ads according to RankScore (the product of bid and AD quality). In advertising systems, advertising quality is generally measured by the estimated click-through rate of advertising.

Ads ranked by RankScore are billed according to the Generalized Second Price.

Therefore, accurate prediction of click-through rate is the premise to guarantee advertising revenue and user experience. Advertising CTR prediction is a typical supervised machine learning problem, whose goal is to accurately predict the probability of click behavior under the premise of given advertising merchant, user and query context. The feature of this supervised learning problem is represented by X, and the target is represented by Y ∈{1, -1} (the click obtained after the advertisement is exposed is 1, otherwise is -1). By collecting online advertising exposure and click logs, we can obtain a large number of annotation samples {(𝑥 I, 𝑦 I)} as training data for supervised learning.

We use the parameter model to fit this probability:

Where, 𝑤 supervised learning problem is an optimization problem that minimizes the objective loss function by searching 𝑤 :

Among them, 𝐿(𝑦, 𝑓(𝑥, 𝑦)) is the loss function of the model, and Negative log-likelihood function is generally used as the loss function in CTR estimation. The optimization problem (Formula 1) is the original click rate estimation problem, and the optimization problem (Formula 2) introduces the regular term 𝑅(𝑤) to control the complexity of the model and prevent the model from over-fitting. In addition, when we select L1 norm as the regular term, we can obtain sparse solutions and reduce the model size, thus reducing the memory requirements of online service loading model and improving the prediction speed of the model. Below we briefly introduce several commonly used CTR estimation models.

Logistic regression model

Logistic regression model is a widely used prediction model of click rate. It is a linear model and the corresponding optimization problem has very good properties. It is an unconstrained convex optimization problem with globally unique optimal solution. It supports large scale features and can converge to the optimal solution quickly by the common echelon method. The logistic regression model has a very good interpretability. We can analyze the importance of each feature and its influence on click rate through the corresponding weight of features.

Logistic regression also has its disadvantages. First, as a linear model, its expressiveness is relatively weak, and a lot of feature engineering work (such as feature combination) is needed to compensate and improve the expressiveness of the model. Secondly, it needs a lot of feature preprocessing, such as feature normalization and discretization.

As a basic model, logistic regression is combined with other models to give full play to its role. For example, the problem of feature discretization and feature combination can be solved through the combination of logistic regression and gradient lifting decision tree, and the support of logistic regression to large-scale features and good optimization problem nature can be fully utilized.

FM model and FFM model of field aware factor factorizer

FM model and FFM model are nonlinear models, which combine features in pairs to improve the expression ability of the model. In addition, both FM and FFM models express and learn features vectorization (Wi, Wi, FJ) to improve the generalization ability of the model. Compared with FM, FFM introduces the concept of domain. In FM, feature I and other feature combinations are represented by the same vector, while in FFM, feature I and feature combinations of different domains are represented by different vectors, which further improves the complexity and expressiveness of the model.

Artificial Neural Network (ANN)

In recent years, the neural network model has a strong revival, and the method represented by deep neural network has surpassed the traditional shallow model in the fields of image recognition, speech recognition and natural language processing, and made a breakthrough progress. Recently, a number of Deep neural network models have emerged in the task of CTR prediction, and achieved obvious results, among which the typical model is Wide & Deep model.

Wide & Deep model consists of Wide and Deep. The Wide part can be likened to the logistic regression model, and the function of relevant features can be well remembered. The Deep part is similar to FM model and FFM model, which both implement vector representation and learning of relevant features. However, the Deep part can express more complex feature interaction and combination relations through complex network structure, providing better generalization ability and expression power.

Gradient method is the basic method of model optimization (optimization problem solution). Formula 3 is the iterative step of formula 1 to solve the CTR estimation optimization problem using the standard gradient method. In the estimation of click rate, due to the large number of training samples (billions, billions), direct application of Formula 3 requires a huge amount of calculation and the iteration speed is limited. Therefore, we generally use Stochastic Gradient Descent (SGD) to solve the CTR estimation problem. In the SGD method (Formula 4), we use the gradients on a small number of samples to approximate the gradient of the overall optimization objective and speed up parameter iteration. Where b is the sample set size, when b=1, we use the gradient of single sample to approximate the gradient value of the overall objective function.

In Meituan, we use the Parameter Server framework to realize model parallelism and data parallelism, so as to solve complex model solving problems involving large-scale training data and features, as shown in Figure 1:

In terms of feature engineering, CTR prediction features are mainly mined and depicted from the three aspects of advertisement, user and query (see Figure 2). Features need to include all aspects that affect CTR, which is an important factor for the success or failure of the model. Feature selection must be based on service scenarios. In the O2O scenario, an important feature that affects click-through rate is the distance between merchants and users.

O2O advertising push

In the O2O scenario, in addition to search and recommendation ads, push ads are also very important. Push advertisement means that the media pushes the right advertisement to the right crowd in the form of message at the right time. The main goal of push advertising is to improve user activity and achieve accurate crowd reach. Meituan has 350 million registered users, but only 30 million daily active users and 100 million annual active users. There are also a large number of users who do not log in to meituan’s App or log in rarely. Push advertisements to these users and guide them to open the App, which helps to improve user activity. On the other hand, push ads achieve accurate delivery through rich crowd orientation. To achieve accurate reach, two points need to be achieved: there is a complete user portrait, user portrait includes attribute label, preference label and behavior label, we can judge the user’s interest in advertising; Intelligent matching technology to accurately target ads to the right users.

The characteristics of push advertising are: active touch, user intention is not clear. In theory, push ads can serve any AD to any user at any time, while search ads can only show ads to users when they search or filter. However, the disadvantage of push ads is that users’ intentions are not clear, while search ads have search terms or clear screening conditions, which are clear user intentions. Therefore, compared with search ads, push ads need more accurate audience targeting.

Audience orientation

There are several commonly used directional methods:

• Time orientation. Time targeting allows brands to target advertising based on consumer behavior, business hours, or even seasonal events or special events. For example, hair salons are open only during the day, and if targeted hours include non-business hours at night, then users can’t call to make an appointment after seeing an AD during non-business hours at night, so there’s no conversion.

• Redirection. It refers to precisely targeting users who have browsed, collected and purchased in merchants according to their historical behaviors, and pushing advertisements back to users to complete the transformation. Often, consumers will remember that you need to redirect without looking at it. Delivering ads based on redirection is a great way to visually remind consumers about a merchant’s product. Consumers may look at this and think, “Oh, I forgot to buy these shoes…” That breadcrumb reminder is often enough to entice them to click and buy. Redirection is the most accurate and has the highest ROI of all.

• Location class orientation. It refers to the real-time location of users (generally cellular information or GPS longitude and latitude) to do some targeting, which helps businesses to reach those consumers who are going to the area of the business, including distance targeting, business circle targeting, etc. This kind of targeted approach is very important when delivering ads on mobile devices. For example, a local hair salon that wants to attract local business can use geolocation targeting to advertise within a specific radius. If the store has a hair salon franchise in the triangle area, it can use the technology for more than one location. Of course, when targeting each region, businesses can adjust their bids according to the business development in the region.

• Population attribute orientation. Demographic attributes include gender, age, income level, marital status, car ownership, and children. Through demographic attribute tags, advertisements can be pushed to relevant consumers, that is, potential buyers. The specific choice of tags mainly depends on what products merchants sell. For example, wedding photography businesses will choose people whose marital status is labeled as “unmarried” to advertise, and nail and eyelash businesses will choose people whose gender is labeled as “female” to advertise. Among these tags, labels such as gender and age are easier to obtain because users provide relevant information when registering. Income level is a label that needs to be estimated. When using demographic targeting, labels should be neither too general nor too segmented. For example, when nail salon beauty and eyelash merchants choose the age label, on the one hand, they cannot choose the age from 0 to 60 years old. This kind of crowd is too extensive, and the low and high age groups may not have a strong demand for nail and eyelash beauty. On the other hand, it is also necessary to prevent too much segmentation of the population. For example, although a more detailed age label may be selected eventually, it should not only be positioned as a specific age. If only people aged 22 are selected, the population coverage may be incomplete. Always keep your target audience in mind, but find a middle ground when targeting it.

• Behavioral orientation. It mines user interests and preferences from user behavior data, and pushes corresponding advertisements. Behavioral data include browsing and clicking of channel, business details page, group order details page, user comments and ratings, etc. Interest preference is generally divided into long-term, short-term and real-time preference. When we mine the long-term preferences of users, we use “behaviors in A period of time”, and need to calculate different weights for behaviors at different times, because the interests of users are dynamic. Three months ago, the preference of users in business circle was A, but now they may change to B when they move to business circle. Sliding window method and time decay method are generally used to measure the weight of behaviors at different times and control the accumulated behaviors within a period of time. Interest preference includes category preference, price preference and business circle preference. Category preference refers to the category of products and services favored by users. For example, whether users prefer Sichuan-Hunan cuisine or Jiangsu and Zhejiang cuisine, hot pot or buffet; Price preference refers to the user’s consumption level, for example, whether the user prefers the price range of 0-20 yuan or 20-40 yuan when ordering takeout.

• Recommended by new customers. In other words, look-alike takes regular customers of advertisers as seed information and combines the big data of advertising platform to find out certain characteristics or laws of regular customers, so as to find potential customers with the same characteristics or laws for advertisers. This approach can ensure accurate directional effect while expanding user coverage. For example, if an advertiser of Sichuan restaurant wants to advertise, the target group can choose the customers who have consumed in other Sichuan restaurant or Hunan restaurant besides the customers who have consumed in this restaurant, because they may have similar tastes.

Users’ behaviors such as searching, browsing, collecting and purchasing on the platform will be recorded, forming user logs. Through the analysis and mining of user logs, user portraits are obtained, including user basic attributes, interest preferences, behavior labels, etc. Advertising targeting is the process of matching advertising and users to find suitable audience groups for each advertisement. After advertising, need to statistics targeted effect, including targeted accuracy and coverage. The operation process of precision targeted advertising is shown in Figure 3.

In order to achieve the matching of advertising and users, first of all, it is necessary to make a preliminary assumption for the audience of advertising, that is, to determine the crowd interested in advertising, and map it to the user portrait label, this step depends on product research and analysis. Then, according to this preliminary hypothesis, the orientation conditions of advertising are determined to match the qualified crowd.

• Representation of a single orientation condition. Each orientation condition is expressed as a
  
  pair. <“Professional”, “Student”> <“Professional”, “Student”>
  ,>
• Representation of combinatorial orientation conditions. The combination of orientation conditions set by advertisers is often very complex, and it is a combination of various orientation conditions, involving intersection, union, inversion and other operations. We use Disjunctive Normal Form (DNF) to store advertising targeting conditions. Here are a few examples of how DNF is expressed.

• • DNF1: (30-year-old male) ∪ (25-year-old female)
• • DNF2: (Cantonese male) ∪ (Peking union)
• • DNF3: (non-male) ∪ (men are located within a 2-kilometer radius of the store in real time) ∪ (likes good food)

First, each DNF can be decomposed into one or more conjoined Normal forms (CNF), DNF1 = C1∪ C2, where C1 = (30-year-old male) and C2 = (25-year-old female). Second, each CNF can be decomposed into one or more conditional intersections. C1 = A1 ∩ A2 in the above example, where A1 = 30 years old and A2 = male.

• Directional condition matching. The directional matching process is shown in Figure 4. A directed request includes a user ID and an advertising ID. First, the user tag is obtained according to the user ID, and the directed packet is obtained according to the advertising ID. The directed packet is resolved and expressed in the form of DNF, and then matched with the user tag.
• Evaluation of directional effect. The directional effect is generally evaluated from both qualitative and quantitative aspects. Quality refers to accuracy, and the main indicators are click through rate and conversion rate. Volume refers to the degree of coverage, and the main indicators are user coverage, advertiser usage, and the proportion of traffic corresponding to targeted methods.

In meituan push ads, redirection has the best click-through rate and conversion rate, but the lowest coverage rate. Location-oriented and demographic attribute tags have a wider range of people and are less effective. The actual use of which kind of orientation, need to see the promotion needs of advertisers, advertisers need to consider the balance of accuracy and coverage.

O2O advertising system tools

Effective tools are an important part of an excellent and efficient advertising ecosystem. In this section, we give a brief introduction from the perspectives of developers, advertisers and operators.

System tools for developers

The tools for developers mainly include three aspects: offline data analysis tools, real-time data analysis tools and online advertising system debugging tools.

Offline data analysis tools support from various dimensions (advertising, advertising type, time, regional strategy, etc.), the algorithm of statistical key indicators of advertising (recall rate, click-through rates, conversion rates, RPS/RPM, CPC, etc.), review of advertising system board and loopholes, help advertising algorithm and the engineering team and looking for potential problems found.

Real-time data analysis tools make up for the shortcomings of offline data analysis in terms of timeliness, helping developers discover data anomalies early, respond to and fix problems faster. Big data processing tools such as Hive, Spark, Elasticsearch, and Druid are behind these analytics. The real-time consumption data analysis tool is shown in Figure 5.

Online advertising system debugging tool is for a single advertiser or a single query and other specific problems of the investigation. Debugging tools make it easy to construct mock requests and view individual service processing details, gather information on each advertising process step (recall, sequencing, creative optimization, etc.), and track and locate problems at each step in the live online environment. In addition to online troubleshooting, debugging tools are also indispensable for verifying the effectiveness of policies and the correctness of algorithms during development. Figure 6 shows the basic interface of an online advertising debugging tool.

Tools for advertisers and operators

Tools for advertisers and operators include advertiser bid estimation and ranking estimation, merchant funnel analysis, account diagnostics and other related tools. Advertiser oriented tools help advertisers better measure and perceive the effectiveness of their ads, give them a better understanding of the market competition, and help them optimize the effectiveness of their ads effectively and proactively. Tools for operators give operators a clearer picture of how advertisers are delivering, which in turn helps them guide and serve advertisers better.

1. Effect funnel analysis tool

As mentioned above, O2O advertisements need to go through multiple processes of click and transformation from online display to in-store consumption. In order to help advertisers optimize the overall delivery effect, we provide an effect funnel analysis tool in the background of promotion. The analysis tool of effect funnel mainly includes three layers of funnel of exposure, visit volume, interest and visit to the store. Meanwhile, corresponding problem diagnosis and optimization suggestions are given, as shown in Figure 7:

2. Promote live tools

Perception of advertising display position and bidding is one of the core needs of advertisers. However, in the O2O advertising scene with personalized intelligent ranking technology system and geographical location restriction attribute, due to user personality tags, geographical location and other reasons, advertisers will not see their advertising exposure in the client, advertisers are difficult to analyze the reasons, and do not know how to optimize the existing advertising.

The live promotion tool provides viewing rankings, mock bids, and diagnostic optimizations. Advertisers can view real-time rankings under certain conditions, such as selected business area, category and geographic location, as well as the average ranking in general after deindividuation. At the same time, the tool will give specific reasons and corresponding hints for the situation that the advertisement display rank is too low or not displayed. Advertisers can adjust the placement Settings according to the prompts, such as suggesting higher prices for lower ranking caused by low bidding. The adjusted new ranking can be viewed in real time through the tool, as shown in Figure 8:

3. Churn order analysis tool

The attrition order analysis tool provides comparative analysis based on the store attrition order record. Lost orders refer to the traffic that users click on merchant A in the last week, but actually place orders to merchants B and C, which is counted as lost orders of A. Analytics tools help merchants analyze the gap between themselves and the merchants that users end up ordering based on their click-to-order behavior data. For example, for hotel merchants, the tool will provide comparison of merchants’ average room price, average score, first picture of merchants and other information, from which advertisers can analyze the reasons for order loss, as shown in Figure 9:

4. Advertising revenue simulator

To attract potential new advertisers, the AD revenue simulator provides merchants with the ability to estimate AD revenue. Based on the store’s historical non-advertising period click conversion rate, traffic in the business circle of the store, and the status of competitors, the tool estimates the new traffic and order volume of the store after the advertising, helps new customers quickly understand the advertising products, and establishes the return on investment expectation. At the same time, through this tool, merchants can easily jump to the Promotion platform for registration and release. In addition, the simulator can also assist sales personnel to estimate the flow of business circles and the number of advertisements that can be carried in business circles, so that sales personnel can explore the market more specifically and improve the success rate of new contracts, as shown in Figure 10:

conclusion

Based on the characteristics of O2O advertising, this chapter introduces the main focus indicators of O2O advertising stakeholders. O2O advertising is one of meituan’s core problems. This chapter focuses on the application of machine learning methods to improve the effectiveness and efficiency of advertising, as well as local scene push advertising. In addition, it also briefly shows the tools related to O2O advertising platform.

Author’s brief introduction

Yiping, who joined Meituan in June 2013, is currently responsible for meituan search advertising algorithm strategy and used to be responsible for Meituan search sorting.

Recruitment information

Meituan advertising platform is fully responsible for the commercial realization of meituan in-store catering, in-store integration (marriage, beauty, leisure and entertainment, learning and training, parent-child, home decoration) and hotel tourism. Based on the real data of hundreds of millions of users, millions of merchants and tens of millions of orders, search advertising is mined to ensure user experience and merchants’ interests while cashing in. Welcome interested students to join the search advertising algorithm group. Please send your resume to zhouyayue#meituan.com

reference

[1] Huang, P.-S., He, X., Gao, J., Deng, L., Acero, A., & Heck, L. (2013). Learning deep structured semantic models for web search using clickthrough data. In Proceedings of the 22nd ACM international conference on Conference on information & knowledge management (pp. 2333-2338). [2] Sutskever, I., Vinyals, O., & Le, Q. V. (2014). Sequence to sequence learning with neural networks. In Advances in neural information processing systems (pp. 3104-3112). [3] Juan, Y., Zhuang, Y., Chin, W.-S., & Lin, C.-J. (2016). Field-aware Factorization Machines for CTR Prediction. In Proceedings of the 10th ACM Conference on Recommender Systems (pp. 43-50). [4] Cheng, H.-T., Koc, L., Harmsen, J., Shaked, T., Chandra, T., Aradhye, H., . others. (2016). Wide & deep learning for recommender systems. In Proceedings of the 1st Workshop on Deep Learning for Recommender Systems (pp. 7-10). [5] Li, M., Zhou, L., Yang, Z., Li, A., Xia, F., Andersen, D. G., & Smola, A. (2013). Parameter server for distributed machine learning. In Big Learning NIPS Workshop (Vol. 1). [6] Edelman, B., Ostrovsky, M., & Schwarz, M. (2005). Internet advertising and the generalized second price auction: Selling billions of dollars worth of keywords.