Push Notifications Evolved: Taobao Knows You Better Than You Know Yourself

11.11 Big Sale for Cloud. Get unbeatable offers with up to 90% off on cloud servers and up to a $300 rebate for all products! Click here to learn more.

By Bei Ling, Kun Qi, and Ming Yi, representing Alibaba’s New Retail and Taobao Technology Department.

Image for post
Image for post

For a long time, Alibaba’s Taobao mobile app in China primarily used push notifications as the main channel for pushing user traffic to platform promotions and events. However, today, as e-commerce platform competition has grown all the more fierce, and user penetration on Taobao has hit record-high saturation, the focus on Taobao has shifted to more refined operation methods and more intelligent delivery capabilities. This new shift is not only for the present but also represents a larger trend for even more upgrades and optimizations in Taobao in the future.

As of this year, the Taobao technology department team have already started to revamp how notifications work on Taobao mobile in terms of all platform interactions, the technical architecture, and delivery algorithms. It is our goal, as the Tabao team, that the push notification feature can come to assume many different roles, including personalized user marketing, user registration and activation, and business diversion in daily platform operations and promotions.

But, you may ask, how have these system upgrades and optimizations played out on the platform? For example, what sort of effect did they have during the last big promotion on June 18? And what are our thoughts and what did we discover? Well, read on to know the answers.

Why General Push Notifications Aren’t Enough

However, nowadays we expect more out of the push notifications. Besides serving as a channel for reaching more shoppers, push notifications should also be able to give pointed and personal recommendations, knowing each and every platform user well, acting like a personal assistant. In reality, push notifications can be implemented based on rich content, a deep understanding of platform users, and personalized operation and distribution capabilities.

The push feature of Taobao mobile classifies push notification messages into marketing messages, product notifications, and instant messages (IMs). Marketing messages are further classified into messages that are personalized, targeted, or general. The following figure shows some examples of push notification messages.

Image for post
Image for post

How Push Notifications Have Evolved on Taobao

With the upgrade implemented in this year, we have set up a personalized online computing engine that is more powerful for content delivery. After performing a lot of AB experiments and testing during big promotions on the Taobao platform, we have successfully switched the main system of push notifications from targeted to personalized delivery, providing personalization in terms of time, content, and frequency.

The following figure shows the differences between the three delivery methods in Taobao’s push notification evolution. In addition to increasing the turn-over efficiency of push notifications, improved delivery methods also provided an enhanced user experience through reducing unwanted notifications and refining the specificity of message content.

Image for post
Image for post

How Push Notifications Work on Taobao

As shown in the following figure, the overall algorithm architecture consists of the delivery matching and traffic shaping modules that are built separately. In addition, a third module named intelligent delivery is also built to determine the delivery time.

Image for post
Image for post

Traffic delivery matching: This module is used to match users and content. We select content and products that are most relevant to users from the material library. For example, if f(U,X) indicates the estimated rate that user U opens the delivered content X, this module selects content that has the highest estimated open rate for subsequent deliveries.

From the system perspective, the entire process is divided into two steps, recalling and sorting. However, this module is different from traditional recommendation and distribution systems in the following aspects:

  1. There are certain constraints and requirements for delivery tasks, including the delivery quantity, target customers, and frequency. If only optimal content is delivered to most active users, this is not the best global solution under these constraints and requirements. Therefore, traffic shaping decision-making is introduced to meet these constraints, which can optimize the global performance.
  2. Different users have different concerns and demands for content at different points in time. This module can effectively increase the content utilization rate and reduce the system load by appropriately estimating the time for triggering a delivery. Therefore, the intelligent delivery module is added to estimate the optimal delivery time for each user.

Traffic shaping decision-making: This module collects user information and content push information to meet certain constraints such as the fatigue and traffic assurance of customers. In addition, this module incorporates these constraints into the algorithm model for optimization.

Assume that M indicates a material, U indicates a user, MU indicates the delivery of the material to the user, and X indicates a delivered content item.

The optimizations and constraints for the final content to be delivered are as follows:

Image for post
Image for post

The h function considers fitting the estimated open rate f(U,X), the sent traffic of U, and the sent traffic of MU. The g function considers optimizing the allocation of the sent traffic of M. Finally, both functions are optimized and learned in a supervised manner to obtain the optimal solution so that the overall traffic allocation can achieve the best performance in terms of traffic, user fatigue, content diversity, and global open rate.

Intelligent delivery: For push notifications, besides message content, the time of the push notification is also very important. Timeliness, or time appropriateness, is required for implementing intelligent delivery. This is, content should be pushed at the time when users are most likely to open the content and will feel least disturbed. With traditional general or targeted delivery, content is delivered at fixed times regardless of the user’s habits.

To correct this shortcoming, we added an intelligent delivery module at the upstream of the delivery system to determine a personalized push time for each user. Before content is delivered, the intelligent delivery module estimates the best push time for each user, and then triggers the content selection and delivery system at the estimated time.

For a user, optimization of the delivery time can be abstracted as the process of estimating the open rate of push messages in different time segments and then selecting the best solution. However, in practice, it is impossible to forcibly solve and traverse all the time points, and traffic cannot be centralized at a single time point, which may cause system errors. The final approach is to select a set (T) of candidate time points based on real-time characteristics, such as the user’s usage or lack thereof on the current day, and then train an estimated model f to select the best point in time as the delivery time for this user on this day.

Image for post
Image for post

Compared with delivery at fixed time and at random time, the open rate of delivery at the time intelligently determined based on scenarios can increase by 10% and 20% respectively, which effectively enhances the utilization of messages. Pushing notification content at the appropriate time also helps reduce disturbance to users and the load pressure of the entire system.

  • Overall results in this phase: After being developed and optimized for three months, the entire system went live just before June 18 and achieved the following improvements in daily delivery:
  • Significant improvement in efficiency:After the system was upgraded, the system doubled its contribution to daily active users (DAUs) on the Taobao mobile app in a few months, starting from the beginning of March, which greatly increased overall user activity on Taobao.
  • Real-time link transformation:After offline delivery was transformed, all computing tasks were migrated to the online engine to implement real-time content matching and delivery. This improved the algorithm performance by over 15% and supported real-time operations. The support for real-time operations is an important, fundamental capability for push notifications to assume a role in big promotions. We will focus on the role of push notifications in big promotions in the next section.

Optimization in disturbance caused by messages:lgorithms are used to predict the value of each message for users. In this way, we can filter certain messages that are less valuable to users and reduce potential disturbance without affecting the overall open rate. With the new optimizations, sent messages were reduced by 40% yet the average open rate was unaffected.

Empowering Bigger Platform Promotions

By transforming our system of push notifications, we worked hard to guarantee incoming traffic for the big promotion through two core capabilities: the real-time delivery and dynamic targeting of push notifications.

The delivery targets were determined based on what was going on during big promotion in real time, with metrics such as the gross merchandise volume (GMV) and the supply and demand of traffic being monitored. It can be difficult for offline algorithms to support such temporary delivery tasks. However, after a series of real-time transformations in the early phase and the survey on temporary demands of the big promotion, the maneuvering strategy function for the big promotion was launched, which is used to implement a series of real-time capabilities, selecting delivery targets and content, and starting a delivery based on temporary or maneuvering demands on the day of the big promotion.

Consider this example, for instance. On June 18, real-time delivery tasks were created to deliver personalized content to target customers who added products to their shopping carts but did not purchase the products. These targeted notifications can help to increase the conversion of potential purchases. We compared the deliveries of manually selected content, randomly selected content, and personalized content selected by algorithms. The following figure shows the experiment scheme.

Image for post
Image for post

The following figure shows how the algorithms improved the delivery efficiency in the three-day-long promotion. Among the three tasks, the open rate achieved by algorithm-based delivery was much higher than that of the open rates achieved by general and random delivery notifications. This therefore proves that the algorithm was of great benefit for reaching the business objectives of the big promotion in terms of efficiency improvements and real-time intervention. In the struggle for traffic during big promotion sales, we hope to reach users in a more targeted and efficient way to help them find more effective information and reduce the interference of useless information.

Image for post
Image for post

Marketing content in big promotions: As you probably already know, a big promotion can be understood as a type of shopping festival for e-commerce users. Therefore, the focus of marketing must be different to traditional shopping festival. Following this, we hope that changes in marketing content can be made in real time and learned dynamically through user behaviors.

During the big promotion on June 18, content was delivered across the entire industry. Based on data analysis, more specifically through comparison between the content delivered by the push notification system during the promotion with content delivered on an average day, we learned that users tend to show more interest in more expensive products during big promotions, whereas on a daily basis they tend towards cheaper items. These changes in user demands can also be captured through real-time algorithm learning and thereby help us to adjust or change push notification deliveries and further improve traffic efficiency during promotions.

Image for post
Image for post

Summary and Outlooks

In the future, we will continue to upgrade and optimize push notifications, mainly in terms of the three following key factors:

  1. Reduce unnecessary disturbance to users and improve notification efficiency: Push notifications can easily become overused in and outside of mobile apps, and it can be difficult to repair damage after the fact. As of now, Taobao mobile push notifications have a lower close rate than other similar apps. However, we still need to continue to optimize the effectiveness of notifications by dynamically reducing unnecessary message notifications.
  2. Support event triggering and provide terminal testing: In addition to marketing notifications, push notifications also need to provide users with events that delight users and notifications that users expect in order to improve user experience and be a real personal assistant for users. Therefore, we will need to continue to supplement the support for event triggering and the real-time computing capability for terminals in the future.
  3. Deeply explore and apply algorithms: As mentioned earlier, the push algorithms of the Taobao mobile app integrate the algorithms used in recommendation and advertising. In the future, we will further explore more algorithms so that we can model user states, push positive and negative feedback more accurately, and make push content more accurate and “useful”.

Original Source

Written by

Follow me to keep abreast with the latest technology news, industry insights, and developer trends.

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store