Alibaba's local life-end intelligent architecture design and technology exploration

Author: Qiu Canqing (Bai Li)

On October 12, 2018, Alibaba Group announced the official establishment of Alibaba Local Life Service Company. Ele.me and Koubeihuishi merged to form the leading local life service platform in China. The mission is to "redefine urban life and make life better"; Koubei focuses on in-store consumption services, Ele.me focuses on home living services, Hummingbird Instant Delivery focuses on instant delivery services, and Keruyun focuses on providing digitally upgraded products and services for merchants, jointly promoting the digitization of the local life market, and making it easy for the world to do things. 's business.

Local life is a very important business map of Alibaba. Based on the business development scenario of Alibaba's local life Ele.me APP, this article elaborates that in the business development process of the Ele.me APP, aiming at the actual business scenario, combining the business through technology, is How to give full play to the advantages of terminal intelligence in terminal-side computing decision-making.

Terminal Intelligence Concept

background

In short, end intelligence is the reasoning operation and upper-layer application of machine learning or deep learning models on the end. Thanks to the current development of hardware devices such as mobile phones, powerful CPUs and GPUs can support some advanced computing operations on the client (in terms of performance and performance). Accuracy can be sacrificed if required), and the computing work can be kept on the client to avoid problems such as data uplink bandwidth overhead and network latency.

Whether intelligence is in the cloud or in the terminal, the problem it solves is the same. How to use a large amount of data as an input source, adjust the algorithm strategy for continuous learning, and infer the final or most suitable conclusion; then end-end intelligence is to put intelligence in The implementation on the device side can standardize the intelligence of the algorithm data. Combined with specific business scenarios, on the device side, the device can collect behavior information and device information in real time, and run artificial intelligence, machine learning and other on-device algorithms for real-time feedback. , adjust algorithms, operations, and product strategies to improve user purchase, experience and other business technical indicators.

characteristic

Compared with cloud-side intelligence, end-end intelligence has significant advantages such as low latency, protecting data privacy, and saving cloud computing resources. The specific breakdown points are as follows:

trend

In the past two years, terminal intelligence has become a popular research and development direction of major Internet companies. Everyone has used many terminal intelligent applications in daily life, such as AI camera for taking pictures on mobile phones, face unlocking FaceID, and various AR in short video apps. Special effects, etc.; within Alibaba, there have been many business explorations, such as Polaroid Taobao, AR makeup try-on, product recommendation, etc., which are typical representatives of the effective combination of terminal intelligence and business.

As part of Alibaba's business map, local life cannot be absent. Based on some brief introductions to terminal intelligence-related concepts in the preface, the following describes the scenes of local life.

local life scene

business attributes

According to reliable data, the overall scale of local life services is roughly estimated to be about 20 trillion, which is a very large market, but the digital penetration rate of this market is still relatively low, only about 10%-15%. The industry attributes of local life are very different from traditional e-commerce. Localization and real-time are the distinctive features of local life services. In essence, local life is actually a service industry. There are very large differences in craftsmanship in the service industry. At the same time, there are many sub-categories in local life. As a result, the service of local life is a non-standard product, which is exactly the same as There is a sharp contrast between the standardization of traditional e-commerce and the easy digitization of commodities.

Ele.me focuses on home-delivery services in Alibaba's local life field. The intelligence on the user side is the only way to better match the demand side with the supply side. Based on the business attributes of home service, the time for users to make decisions is very short. Basically, the peak period is during the lunch and evening meal time periods. In a limited time period, how to analyze the current needs of users and analyze the current needs of users. In order to better match demand with our supply, terminal intelligence can play a very good role in this field.

Technology Architecture

The local life terminal includes multiple apps such as Ele.me, Koubei, Hummingbird, Merchant Edition, Xuanyuan, etc. There will be different application scenarios of terminal intelligence on different apps; based on the consistency of the local life technology architecture, we abstract a The terminal intelligence adaptation layer of local life is customized based on the business attributes of local life. At the same time, combined with the rich middleware and platforms in the group, iterative development of terminal intelligence-related applications can be carried out quickly.

The current smart technology architecture of the local living terminal is as follows:

Based on the general group middleware and end intelligence SDK, the end foundation and end intelligence foundation are abstracted. The end foundation includes business buried point library UT, switch configuration Orange, high-speed data transmission Highway, real-time monitoring Answer, database operation SQL, The network library Mtop, etc., the end intelligence foundation includes the deep learning end-side reasoning engine MNN, end-side user behavior data BehaviX, end-end intelligence running environment Walle, etc.

According to the technical characteristics and business characteristics of local life, the adaptation layer ALSCAdapter of local life is designed. In the adaptation layer, we have added many features required by business, such as application life cycle, automatic buried point, etc., and combined with local The monitoring system of life creates an intelligent real-time monitoring system. APIs have been opened for different businesses. In order to solve the problem of low efficiency in development, debugging and testing, we have developed an end-side Debug platform on the local life side, and jointly built it with the Group's MNN workbench to improve the efficiency of algorithms and testing. .

Technology exploration

Based on the unified terminal intelligent technology architecture of local life, we can develop and iterate better and faster in business, and also help other APPs in local life to quickly access and get started. On the Ele.me APP, we mainly made some technical explorations in several fields such as user characteristics, intelligent recommendation, and intelligent reach, and achieved certain business results. The following is an introduction one by one.

User characteristics

Richer and more real-time user characteristics are the key to improving the accuracy of algorithmic decision-making. Based on the ability of terminal intelligence, the key behavior data of users is stored on the terminal, such as user entry and exit data, store entry and exit data, additional purchase order data, Browsing behavior data, etc., can collect very rich metadata on the terminal side.

Based on metadata, users’ behavioral characteristics are extracted in real time, and the latest user behavior characteristics are maintained in real time. After returning to igraph, the delay is about 2s, so that the algorithm model in the cloud can obtain the user’s behavior characteristics in advance, and at the same time, the characteristic data is locally persisted. Other end models can also directly obtain relevant feature data on the end, so as to extract it once and use it in multiple parties. In order to standardize the process, we specify the production rules of end features, and divide the function of the feature data table on the end to improve the efficiency of obtaining local features.

The current overall process of feature extraction on the end is as follows:

By continuously expanding the types of features and establishing an on-device feature center, feature metadata is provided for subsequent device-side model calculation and cloud model improvement, allowing business algorithms to focus on the implementation of business details without the need to extract their own features for each business, saving development time It can also avoid the waste of resources on the end.

Intelligent Recommendation

Search recommendation is a business area where end intelligence can play an important role. For recommendation feeds, due to paging loading, there are 20-30 store data in one page of data. The time for algorithm to intervene is very small, and each intervention can only be done in The user browses to the next page, and the cloud recommendation algorithm can make intervention decisions on the recall results when the client triggers a page-turning request. The overall process is very long. User->client->engineering->algorithm, the data on which the algorithm decision depends The delay of the return link is at the minute level, which is completely unable to meet the needs of real-time recommendation. The recommended results may not have much correlation with the user's previous behavior.

In the Ele.me APP, since the user's usage time is not very long, the algorithm needs to make multiple decisions when the user browses the feeds and give the user the optimal solution.

The terminal intelligence-based recommendation rearrangement scheme can rearrange strategies for stores that users have not browsed, and make weighted adjustments according to users’ previous behaviors. Trigger the algorithm so that the algorithm model on the terminal can intervene in real time to achieve the algorithm goal; the logic of the overall recommendation rearrangement is as follows:

The timing of the rearrangement trigger is agreed with the algorithm, and some boundary triggers such as a small number to be rearranged are excluded. The algorithm model uses cloud features and on-device features to make decisions together, and conducts AB experiments through models with different business goals, such as gmv goals, ipv Goals, etc.; currently, it has been applied in the Ele.me homepage recommendation feeds and the food channel feeds, and achieved good business results; at the same time, in order to expand the number of candidate pools and give more decision-making space to the algorithm on the side, we also did Smart refresh of the terminal to replenish the store candidate pool in time; interactive recommendation and smart weather vane are also under planning and development.

Smart touch

In the field of user growth, how to reach users efficiently and accurately is the key to user growth. Push is the channel that can reach the most users, whether it is inside or outside the terminal. Before intelligent touch has been proposed, only batch push and circle push can be configured to users. The push configured in this way is too rigid, resulting in a very low click-through rate of the overall push, and will disturb many users. This increases the chances of users turning off push permissions. Operation students need more accurate timing and richer contact methods to achieve business goals.

Intelligent touch based on terminal intelligence, through the deployment of the terminal model on the client, through the business rules and operation configuration to intelligently identify the contact points that may be the inflection point of user behavior, combined with the characteristics of the terminal cloud, interact with the push center, and the cloud algorithm horse racing mechanism to obtain the optimal push the configuration, and finally reach the user. Based on the intelligent touch in the terminal, since the inflection point of the user's behavior is identified, the push that occurs at this time will not disturb the user, but will become an assistant to promote the user's order conversion. Based on the intelligent touch outside the terminal, it mainly collects the characteristics of the user's behavior on the front and back, reports it in real time, and identifies the user's back-end intention for back-end recall. The overall smart touch solution is as follows:

Through the production of smart touch points, the operation students have more and better opportunities to reach users. Combined with the customized configuration of the operation platform, they can produce thousands of people and thousands of faces. Through the launch of Smart Touch, the click-through rate of push in the terminal has been greatly improved, and the recall of the back-end terminal also has a very good user retention effect, helping business operations to achieve the goal of user growth.

Other business

In addition to the above-mentioned business and technology exploration, we have also made other attempts at end-to-end intelligence, such as Hummingbird's Blue Storm project. By carrying out end-to-end pre-recognition projects, we help riders to improve their photo-taking experience and save riders' operations during the rider's random inspection process. At the same time, it also reduces the cost of manual review, improves the service quality and brand image of Ele.me, and ensures the safety and brand image of the food delivery process.

future outlook

In the future, in the environment where the performance of mobile devices is continuously enhanced and the network infrastructure is continuously improved, the room for end-to-end intelligence will become larger and larger. In the field of local life, the continuous promotion of real-time search recommendation and operation intelligence is the future development of end-end technology. For the way forward, I would like to thank the Group's end-end intelligent technical team for their support and help, and also thank the close cooperation of various related parties such as local life products technology algorithms, so that end-end intelligence will continue to blossom in local life scenarios.

In the future, we also need to better precipitate the capabilities of end-to-end intelligence into infrastructure to help various businesses access quickly, make algorithm experiments more flexible, data recovery faster, and continue to explore more new business scenarios. We need more like-minded people who are willing to devote themselves to the field of local life, work together to grow together, and use technology to help business shine! (ps: People of insight who are interested in local life are welcome to join, resumes can be sent to [e-mail] )

Follow [Alibaba Mobile Technology] WeChat public account, 3 mobile technology practice & dry goods every week for you to think about!