Open screen ads = make users wait? How does Xiaohongshu balance user experience and advertising effect?

As an important part of Xiaohongshu's brand advertising, open-screen advertising has become a powerful booster for many brand customers to reach target users and realize brand marketing appeals. The placement strategy of Xiaohongshu’s open-screen advertisements must not only meet customers’ demands for quantity and effectiveness, but also take into account the impact on user experience. As a result, we have developed a "traffic optimization + dynamic decision-making" solution . Starting from the formal optimization of screen-open advertising traffic, we derived the optimal allocation strategy and achieved online traffic allocation based on feedback adjustment. At the same time, in order to shorten the screen-opening scene as much as possible To reduce the user's waiting time, we designed a dynamic decision-making mechanism to optimize traffic while meeting user experience requirements.

Open-screen ads will be displayed in full screen when users open the Xiaohongshu APP. The immersive visual experience brought by full-screen exposure can effectively help brands and products enhance user awareness. With rich interactive styles and landing page types, Xiaohongshu's open-screen advertising has become a powerful booster for many brands to quickly reach target users during new product launches, brand activities, etc.

At present, Xiaohongshu's open-screen advertisements mainly support sales through contracts . That is, when customers purchase traffic, they confirm the scheduled exposure amount with the platform in advance, and the platform automatically places it on the day of release and ensures that the exposure amount meets expectations.

In recent years, under the background of combining product and effect, in addition to ensuring quantity, customers who place open-screen advertisements also hope that the platform can optimize the effect of advertising . In actual business scenarios, customers mainly evaluate advertising performance based on CTR. In the past, Xiaohongshu's opening-screen advertising focused more on ensuring order quantity, without fully considering how to optimize the delivery effect. We hope to optimize the opening-screen advertising by upgrading the advertising strategy to traffic optimization while meeting the order quantity guarantee. CTR.

Next, we will introduce Xiaohongshu’s opening-screen advertising strategy in two parts:

● The first part introduces how we can meet customers' demands for the volume and effectiveness of open-screen advertising through traffic optimization under ideal circumstances. We started from the definition and formalization of the problem, derived the optimal allocation strategy for traffic optimization, and implemented online traffic allocation based on feedback adjustment;

● The second part introduces the upgrade of decision-making mechanism . When the project was implemented, due to Xiaohongshu's high requirements for user experience, the APP startup time was compressed to a very short time, making it difficult to make real-time advertising decisions. We designed an asynchronous decision-making solution, and based on this, we further implemented open-screen advertising. The dynamic loading mechanism implements screen opening traffic optimization on the premise of meeting user experience requirements .

The traffic optimization of Xiaohongshu's open-screen advertisements is to optimize the advertising CTR as much as possible while ensuring quantity. This is a typical traffic allocation problem. A common approach in the industry to address this type of problem is to formally model the allocation problem, transform the problem through dual solution, and solve for the optimal traffic allocation formula. Then, based on the optimal allocation formula, feedback adjustment is used to achieve optimal traffic flow. Following this idea, we first introduce the formal modeling and solution of the problem.

Problem definition and solution

We define the open-screen full-day traffic set as $\{pv_i \ | \ i = 1,2,3,\dots,n\}$ , the advertising order set as , and the volume guarantee target of $\{order_j \ | \ j = 1,2,3,\dots ,m\}$ each order as , which represents the result of traffic optimization, which means that the th traffic is allocated to the th order ; pctr on each traffic; considering that in some specific business scenarios, some orders particularly need to optimize CTR, we introduced click value weights to represent the importance of CTR optimization for different orders. According to the above parameter definitions, the traffic optimization problem can be modeled as: $order_j$ $d_j$ $x_{i,j} \in [0,1]$ $x_{i,j} = 1$ $i$ $j$ $pctr_{i,j}$ $j$ $i$

Formula (1) indicates that each order needs to satisfy the quantity guarantee constraint, and formula (2) indicates that each traffic can only be allocated to at most one order.

Considering that during actual delivery, the quantity may not be guaranteed due to excessive sales, and the above problem has no solution, so the shortage penalty weight of each order is introduced to $w_j$ convert the problem into:

The above problem is a typical linear programming problem. Through dual transformation and complementary relaxation theorem derivation, the optimal ranking score (rank score) calculation formula can be obtained: , where is the dual variable of the order-th order quantity $\begin{array}{l l} rank\_score_{i,j} = pctr_{i,j} \cdot v_j + w_j - \alpha_j \end{array}$ preservation $\alpha_j$ constraint $j$ .

The traffic optimization strategy corresponding to this optimal formula is: all orders on the same traffic calculate their respective rank scores based on the above formula. If the rank score of an order is greater than 0, the order with the largest rank score will win the traffic; if the rank of all orders If the scores are all less than 0, then there is no order to win the traffic, and the user will not see the opening screen advertisement when launching the APP this time.

Online allocation strategy

Through problem formalization and derivation, we obtained the optimal traffic distribution strategy. Under the premise that the all-day traffic set is known, the optimization of each order can be obtained directly through LP solution $\alpha_j$ , and then the rank score of each order on all traffic is calculated to obtain the traffic optimization result. During the actual implementation, it is considered that the advertising traffic on the screen will fluctuate to a certain extent every day, and the online traffic is difficult to predict, so it cannot be directly solved. We refer to common solutions in the industry, transform the LP problem solution into a parameter control problem, and use feedback adjustment to dynamically update the parameters of each order $\alpha_j$ .

According to the formula $\begin{array}{l l} rank\_score_{i,j} = pctr_{i,j} + w_j - \alpha_j \end{array}$ and the corresponding traffic distribution strategy, when the rank score is less than 0, the order cannot compete for the corresponding traffic, that is, each order can only compete for the traffic that meets the requirements. The parameters directly $\begin{array}{l l} pctr_{i,j} + w_j \geq \alpha_j \end{array}$ determine $\alpha_j$ the order $order_j$ 's competing traffic range. Further analyze the impact of parameters on order delivery: for any order, if it is adjusted down when actually allocating traffic, the traffic $\alpha_j$ satisfied by the order $\begin{array}{l l} pctr_{i,j} + w_j \geq \alpha_j \end{array}$ will increase, the participating traffic will increase, and the overall rank score will increase, and the amount that can be competed will increase. The increase in traffic will speed up the delivery speed; otherwise, the delivery speed will slow down. The idea of feedback adjustment is to dynamically adjust the delivery rate of each order according to the delivery progress of each order $\alpha_j$ , so as to optimize the delivery effect while achieving the volume target.

In terms of specific solutions, the algorithm strategy will obtain the cumulative exposure of the open-screen advertisement on the day of investment every once in a while (such as 5 minutes). At the same time, it combines the volume maintenance targets of different orders and the distribution of market traffic to obtain the expected exposure of each order. Exposure. For each order, feedback is updated based on the respective expected exposure and actual exposure $\alpha_j$ . Judging from the actual delivery effect, this method can better optimize the CTR of open-screen ads while achieving the volume target.

High user experience requirements

For a community platform like Xiaohongshu, the majority of users are the foundation for the sustainable development of the platform. Xiaohongshu has always attached great importance to user experience . We don’t want users to wait too long when launching the APP because of open-screen ads, resulting in poor user experience or even user loss. Therefore, the APP startup time is compressed very short.

So how did Xiaohongshu’s previous opening screen ads meet the startup duration requirements?

In the early days, Xiaohongshu’s on-screen advertisements used a mechanism of loading materials + decision-making in advance to avoid the impact of the APP startup time limit. Since the opening screen mainly supports contract sales, the advertisements to be placed on that day have been determined before the start of each day. In the early days, on the one hand, considering that the number of advertisements was relatively small, all materials could be loaded onto users’ mobile phones in advance; on the other hand, each user’s exposure content could be calculated in advance based on the advertising orders that had been sold, and the calculation results could be sent to On the user's mobile phone, mark the advertising content that each user sees every time that day.

Although this method of deciding the advertising exposure results in advance will not be affected by the APP startup time limit, it will not dynamically adjust the advertising content that users see based on the delivery status of each order, and cannot adapt to the feedback control of traffic optimization. It is difficult to fully unleash the effect of flow optimization. In order to adapt to traffic optimization, the common practice in the industry is to make real-time decisions on advertising content in the opening scene. That is, every time a user opens the APP, the client requests the advertising system to make online decisions in real time, and returns the results to the client to complete the advertisement. exposure.

However, in the opening scene of Xiaohongshu, due to the short start-up time of the APP, the limited time in many cases cannot support us to complete the entire process of "advertising request → online decision-making → result delivery" , so it cannot be like other platforms. Enable real-time decision-making. In order to solve this problem, we designed a new decision-making mechanism to achieve decision-making on advertising content and release the effect of traffic optimization on the premise of meeting the APP startup time requirements.

Asynchronous decision-making mechanism

As mentioned above, many times the execution time of APP startup is not enough to decide the content of "this" screen exposure, so we changed our thinking and still maintained the mechanism of loading all materials in advance, but adjusted the way of advertising decision-making. We assume that the interests of the vast majority of users will not change significantly between two consecutive screen exposures. In this way, the decision-making mechanism can be adjusted from the decision-making "this time" without losing the advertising effect as much as possible. Content that will be exposed “next time” for decision-making. Through asynchronous interaction, the result returned by each ad request is used for the next user's exposure when the screen is opened. This way, it is not limited by the APP startup time, and optimizes the traffic of open-screen ads. The specific logic of this asynchronous decision-making mechanism can be referred to the following figure:

After the user completes the opening screen advertisement exposure, the APP will initiate an ad request. The advertising system will determine the display content of the user's next opening screen advertisement based on traffic optimization and return it. The client will cache the returned results locally; the next time the user opens the APP At this time, the local cache of the APP is read, the corresponding advertisement is displayed to the user, and the next advertisement content is continued to be requested after the exposure is completed.

Dynamic decision-making mechanism

The asynchronous decision-making mechanism is not perfect. Under the asynchronous mechanism, there is a natural delay between ad request and exposure, which causes a decision-making delay in traffic optimization and affects the effect of optimization. In order to improve the timeliness of open-screen advertising decisions, on the one hand, we have appropriately increased the request timing of asynchronous decisions and shortened the "request → exposure" interval in traffic optimization as much as possible; on the other hand, we have also made some APPs that take longer to start In the traffic, real-time requests are implemented to further improve the timeliness of advertising decisions.

The specific plan is that we divide the advertising traffic into hot start and cold start according to the type of APP startup. Hot start mainly refers to the opening screen advertisement exposed when the APP switches from the background to the foreground. At this time, the APP startup process is very short. We use the method of directly reading the existing asynchronous decision results to determine the ads to be displayed this time; cold start is the opening screen advertisement when the APP starts from a state without background. Since it involves APP initialization and other processes, the startup time is long. You can use a real-time request solution to execute the entire process of "advertising request → online decision → result delivery". Through "hot start + asynchronous decision-making" and "cold start + real-time decision-making", we have realized the dynamic decision-making mechanism of Xiaohongshu's open-screen ads, and based on this, we have implemented traffic optimization for open-screen ads.

Online practice has confirmed the effectiveness of our solution. Compared with the early combination strategy of volume priority + early decision-making, the new traffic priority + dynamic decision-making solution can significantly improve the performance of open-screen contract advertising while meeting the volume requirements. The CTR of the advertisement ; and unlike the common real-time decision-making scheme in the industry, the dynamic decision-making scheme has basically no impact on the APP startup time, shortening the user's waiting time as much as possible, and ensuring the user experience of opening the Xiaohongshu screen.

In order to meet customers' demands for volume guarantee and delivery effect, and taking into account the high requirements for user experience in Xiaohongshu's screen opening scene, we implemented a complete solution of "traffic optimization + dynamic decision-making" mechanism to achieve business goals. Future iterations are mainly divided into the following directions:

1) In terms of parameter control, the current parameter control for traffic optimization adopts a strategy of independent updates for different orders. The parameter updates of each order affect each other. In the future, more complex control strategies based on RL or MPC will be considered to solve the above problems;

2) In terms of delay decision-making, under the dynamic decision-making mechanism, the delay from ad request to exposure is common, which in turn affects the effect of delivery. In the future, we will consider introducing the prediction of user's subsequent behavior in online decision-making for optimization;

3) In terms of material loading, the materials for open-screen advertisements are still loaded in full in advance. As the number of advertising materials increases, it will affect system performance and increase the storage burden on the client. We will also explore a real-time material loading mechanism to reduce unnecessary Material delivery.

Crayon (Han Zhongxing) Xiaohongshu Information Flow Advertising Algorithm Department

Xiaohongshu mechanism strategy algorithm engineer, graduated from the University of Chinese Academy of Sciences with a master's degree. He has rich experience in mechanism strategies of effect advertising and brand advertising. He is currently mainly engaged in work related to brand advertising mechanism strategies.

Xu Feng (Tian Zengguang) Xiaohongshu Business Technology Engineering Department

Xiaohongshu commercial back-end development engineer, graduated from Beijing University of Posts and Telecommunications with a master's degree. He has been engaged in engine architecture construction and product capability building in the direction of performance advertising and brand advertising. Now he is focusing on the continuous iteration of commercial brand advertising.

Qingshan (Xing Zhizhuang) Xiaohongshu Information Flow Advertising Algorithm Department

Xiaohongshu mechanism strategy algorithm engineer, graduated from Beijing University of Posts and Telecommunications with a master's degree. He has been engaged in related work in the direction of brand advertising and customer mechanism. Now he is focusing on related research and exploration in the direction of brand advertising and traffic mechanism.

Open screen ads = make users wait? How does Xiaohongshu balance user experience and advertising effect?

Guess you like