Technical dry goods | Xianyu: What kind of past and present needs to be experienced for an excellent Push platform

Introduction to mPaaS message push service, quickly integrates Push channels from multiple vendors, effectively improves user retention and improves user experience.

Editor’s Recommendation:

Click here to learn about the mPaaS message push service, quickly integrate Push channels from multiple vendors, effectively improve user retention and improve user experience.

_The following article comes from Xianyu Technology, author Jianxin_

For Xianyu users, because Xianyu has only one product in inventory, the timeliness of the product is very strong. Therefore, when the price of a new or browsed product is reduced by the seller that the user follows, or the platform finds a batch of cost-effective products for the user, the user Hope to be notified as soon as possible. Push has become an important link between users and the Xianyu platform.

This article will introduce the development process of Xianyu Push from the offline manual release 1.0 version to the intelligent and personalized 2.0 version from the perspective of technical students, and explain in detail the problems encountered and the selection of technical solutions, in order to bring readers some thoughts and ideas. Ideas to solve similar problems.

idle fish Push1.0

When Xianyu is all in wireless, the platform needs to push high-quality content related to users to users, so that users can quickly find the products they want to buy and the content of interest. The platform urgently needs a Push productization solution to ensure that high-quality content reaches users in the form of Push and enhance user experience. Based on this premise, the main ideas of the Xianyu Push1.0 solution are as follows:

1. Calculate the list of Push users

a. Calculate high-quality Push scenarios that are strongly related to users, and get a list of users according to the scenarios

b. The vertical business circle selects the target group of the business based on user portraits and other conditions

2. Filter the list of scenes that each user can send based on scene fatigue

1. Globally select the best of each user's scene list, and select the scene with the highest click rate as the target scene

The Xianyu Push1.0 solution is simple and clear, the process is clear, and the offline process is convenient for monitoring alarms and troubleshooting, meeting the business needs at the time, and running stable after going online. In a long period of time, the core structure and process of the 1.0 program have not changed much. However, with the development of business, some of the drawbacks of Xianyu Push1.0 solution began to be exposed, including

• Push user list calculation is not real-time enough
• The message card style is not rich enough
• Single Push trigger timing
• There are fewer Push scenarios

These problems eventually led to the inability to continue to increase Push's click-through rate, which touched the ceiling of the 1.0 solution. In order to solve these problems, we optimized and upgraded Xianyu Push system in several aspects, and finally reconstructed Xianyu Push system.

idle fish Push1.1

In order to provide Push users with a better user experience and enrich user Push scenarios, we give priority to optimizing and extending the existing Xianyu Push solution in terms of message style, trigger timing and user scenarios. The optimization items are mainly divided into message feeds stream upgrade, Push time personalization, real-time Push, etc.

News feeds stream upgrade

Push will be deposited in the message section of the client, and the message section is also one of the most frequently viewed sections after users enter Xianyu. Initially, the message style only supports text messages and image messages. The problem with this type of message style is that it is effective for users. Information exposure is small, and the message style is single. To this end, we have upgraded the message display format, displayed messages through feeds, increased effective information exposure, optimized message styles, and created user minds in the message section. After the feeds stream is upgraded and launched, the effect is obvious, because the content that users are interested in reveals more than that, and the relative increase in UV click rate and user retention is relatively large.

The first one is feeds stream messages, followed by picture messages. Relatively speaking, feeds stream messages can reveal more effective information.

Push time personalization

The Xianyu Push1.0 solution mainly supports regular batch Push, and the actual operation situation is to send Push to target users in batches regularly. Push trigger timing is relatively simple, and the traffic is more concentrated by humans, which increases the system stability risk; in addition, the unified trigger timing does not apply to all users, and it may disturb some users.

In response to this situation, we optimized the Push trigger timing, and the algorithm calculates and predicts the trigger timing of each user based on user behavior. The algorithm divides users relatively evenly into one day, and pushes to users during the relatively active time period of users, reduces harassment of active users, and makes the user group stratification reached by Push more reasonable and healthy.

Real-time Push

The Xianyu Push1.0 solution mainly covers user-related offline scenes, which does not cover enough scenes generated by users' real-time behaviors. Moreover, compared with offline scenes, such scenes have higher real-time performance and are relatively more important to users. In response to this problem, we have increased the coverage of real-time scenes, abstracted user behavior into a relational model, and used IFTTT as the overall trigger mechanism of the system. When the user's behavior on one side of the relationship changes and the other side is touched, this type of scenario is more real-time and strongly related to the user, enhancing the richness of the user's Push scene and enhancing the user's stickiness. After the real-time Push scene is online, the Push click rate is more than doubled compared to the offline scene. For specific technical details, please refer to " IFTTT ".

The above is our function optimization and enhancement for Xianyu Push1.0. These capabilities have also been expanded to support more scenarios and services, and finally combined into Xianyu Push1.1 version.

The life of the idle fish Push

The overall launch of Xianyu Push1.1 has greatly improved the richness of users' Push scenes and user experience. With the deepening of understanding of Push and users, we found that there is still room for optimization and improvement, including:

• The platform's perspective is not enough, and the existing optimization is more point-to-point. It is necessary to connect these points into a line from the perspective of the Xianyu Push platform to form a resultant force, resulting in an effect of 1+1>2
• The essence of the existing Xianyu Push process is still offline computing, and the algorithm cannot perform more real-time personalization and global optimization, which will have a certain impact on the user experience
• The scene configuration is not flexible enough, and the cost of adding new scenes is high, which restricts the progress of enriching user Push scenes

For these reasons, we finally reconstructed and upgraded the Xianyu Push system to build the Xianyu Push real-time intelligent delivery platform Hermes. Hermes is taken from Greek mythology. He is smart (intelligent), agile (fast), versatile (multiple reach), and fits the mission vision of Xianyu Push's real-time intelligent delivery platform best.

logical architecture

The Hermes architecture is completely different from Xianyu Push1.0. With real-time as the goal, real-time or quasi-real-time is achieved in key links such as scene material preparation, global algorithm tuning, and Push sending, so as to improve the timeliness of Push content; in addition, from a platform perspective, Hermes is divided into a configuration center, a matching center, and a task center. Each subsystem defines the interactive data protocol, and there is no strong dependence on each other. The functions of the three subsystems are:

• Configuration Center

  The configuration center is responsible for maintaining the core data model of the platform, providing the business side with page operation configuration Push scenarios and materials, reducing the business side's access cost; and synchronizing the configuration data to the algorithm model in an offline full amount and real-time incremental mode as a matching basis.

• Match center

  The matching center is also called the algorithm selection center. The matching center is responsible for training the algorithm selection model, personalizes screening for each user according to the scene and material configuration, sorts according to the historical click rate data of each material, and recalls the user most likely to feel the most likely to be recalled according to the user's recent behavior Interest materials and personalized content.

• Task center

  The task center is responsible for the timing and actual reach of Push triggers. The core of the task center supports timing triggers, real-time triggers, and time personalized triggers. The purpose is to close the Push trigger methods, choose different trigger methods for different businesses and scenarios, and help businesses achieve business goals . In addition, it is to close the reach to facilitate the platform to arrange the reach plan, including the trigger time and the sending level, to ensure that the business goal is reached while not causing excessive instantaneous pressure on Hermes and the downstream system of the business.

Business effect

After the Hermes platform went online, the effect is very obvious, mainly as follows:

• Push click-through rate increased to double digits
• Directly double the coverage of user scenarios
• The DAU activated by Push also exceeds the highest level in history

summary

This article introduces the whole process of Xianyu Push from the 1.0 version of offline computing in the previous life, to the 1.1 version with multiple optimizations, and finally evolved into the real-time intelligent delivery platform of this world. In fact, each stage of Xianyu Push fits the business development at that time. Yes, but the infinite pursuit of user experience finally produced Hermes, a real-time intelligent delivery platform for Xianyu Push. I hope this method can help readers understand the business background and technical solution selection considerations of Xianyu Push.

Author of this article: )

Click here learn more about mPaaS news push.

Copyright Notice: content of this article is contributed spontaneously by Alibaba Cloud real-name registered users. The copyright belongs to the original author. The Alibaba Cloud Developer Community does not own its copyright and does not assume corresponding legal responsibilities. For specific rules, please refer to the "Alibaba Cloud Developer Community User Service Agreement" and the "Alibaba Cloud Developer Community Intellectual Property Protection Guidelines". If you find suspected plagiarism in this community, fill in the infringement complaint form to report it. Once verified, the community will immediately delete the suspected infringing content.