Shulou(Shulou.com)06/03 Report--

With the continuous development of offline scene layout and the continuous advancement of online technology, an era that truly belongs to new retail has come.

After walking through Guangzhou, Chengdu, Beijing, Shenzhen and other four cities, TechDay, a push technology salon, ended perfectly in Shanghai. The technology experts from Jingdong Home, Tui, Yikatong and Pinlan analyzed and demonstrated the new technology in the new retail era at the scene of Shanghai Station.

"Practice and Application of User Portrait in Big Data Era" Ye Zhengjun Pushes Big Data Analyst

User profile, that is, labeling of user information, is essentially labeling of data. There are three common user profile systems: structured, unstructured and semi-structured. Unstructured systems have no distinct hierarchy and are more independent. Semi-structured hierarchies have a hierarchy concept, but no overly strict dependencies. In the e-commerce industry, more enterprises will choose semi-structured user portrait system for application. Take a simple three-level structured tag as an example. The first-level tag has basic attributes and interest preferences, and can be extended to the second-level tag and the third-level tag.

In the aspect of portrait construction, developers can refer to some common practices, such as label system design, basic data collection and multi-data source data fusion, realizing unified user identification, constructing user portrait feature layer, portrait label rule + algorithm modeling, algorithmic labeling and portrait quality monitoring for all users, etc.

The overall process of pushing user portraits can be divided into three parts. The first is basic data processing. Basic data includes device information, online APP preference data and offline scene data. Second, the image intermediate data processing. The processing results include online APP preference features and offline scene features. Third, portrait information table. There should be four kinds of information in the table: basic equipment attributes, basic user portraits, user interest portraits and other user portraits. At the same time, the construction of user portraits requires the joint participation of technical and business personnel to avoid formalized user portraits.

In the process of constructing user portraits, the main technologies used by each push are data storage, real-time computing, machine learning and deep learning.

Applications for user profiling include: precision marketing, user analytics, data mining, service products, industry reports, and user research.

In view of the demand of APP operators for user portraits in the new retail era, relying on the accumulation of years of push services and powerful big data capability, the company launched the user portrait product "Personal Image," providing APP developers with rich user portrait data and real-time scene recognition capability.

At the same time, the unique cold, hot and warm data labels can also effectively analyze the online and offline behaviors of users, mine user characteristics, and help APP operators screen target groups. At the same time, a push can also provide customized labels for APP to meet the needs of APP operators in user digital management.

In practice, the integration of customized labels also has certain difficulties. Each push will combine the data of both parties, model and analyze them, and output customized labels. In general, each push not only has a generic label dimension, but also has the ability to customize the label output.

Take two typical applications of user portraits in a push business: first, accurate recommendation, APP operators can display different content to users through gender, age level, interests and hobbies provided by individual images, so as to achieve accurate operation and thousands of people.

Second, user clustering, processing user data provided by customers, completing user portraits, and finally performing user clustering analysis.

"Machine Learning Practice in Real-Time Logistics Scenarios" Zhuang Xuekun Dada-Jingdong Home Logistics Algorithm Team Leader

As a new retail "water, electricity and coal," real-time logistics is at the core of the new retail mode, solving the problem of commodity distribution efficiency. As the leading platform of domestic instant logistics, Dada-Jingdong Home has carried out a lot of technical exploration and accumulation in this respect. Compared with traditional logistics mode, distribution in real-time logistics scenario has higher complexity, which is manifested as follows: 1. Diversification of order types;2. Higher timeliness requirements;3. The transportation capacity of delivery knight is difficult to control;4. The destination of delivery is complex and diverse. The problems and challenges existing in the form of real-time logistics can also be summarized into four parts: highly dynamic logistics orders, dynamic distribution costs, the need to balance fairness and efficiency in order distribution, and the management of knights 'freedom to grab orders.

Nowadays, new algorithmic models emerge one after another, and the degree of freedom that algorithms can choose is relatively high. But in practice, the data determines the upper bound of machine learning, and the algorithm simply tries to approximate that upper bound as much as possible. A successful and practical algorithm system must attach great importance to feature engineering. The premise for developers to develop an excellent algorithm system is to obtain high-quality data with accurate characteristics. Dada has accumulated a large and accurate set of distribution scene features by combining its own distribution scenes.

There is no best algorithm, only the most appropriate algorithm for a given scenario. After obtaining the feature data, the machine learning system applied to the real-time logistics scenario can be divided into four layers: the basic data layer, the feature engineering layer, the algorithm model layer, and the business application layer. The development team also needs to select different algorithm models according to the application scenarios of the business. For example, linear models are good at dealing with high-dimensional microscopic features, while nonlinear models are good at dealing with low-dimensional macroscopic features. In path planning and scheduling, traditional operational research models such as dynamic planning may be more appropriate.

Finally, a more general AI can be composed of two parts, the first part is Deep Learning, which solves end-to-end learning problems, and the second part is called Reinforcement Learning, which allows for a more general learning architecture. If these two parts are combined, they can become a very general learning algorithm. In the design of Dada intelligent supply and demand regulation system, the idea of AlphaGo is used for reference, and the advantages of this mode are fully exerted, so that the effect of regulation efficiency and the degree of automation are greatly improved at the same time.

"Enhancing New Retail Service Chain with Intelligent Dialogue Robot" Kong Xiaoquan Geely Group Ecarx Algorithm Expert

Different from the previous retail mode, there is no supermarket feedback and intermediate chain in the process of new retail, and enterprises need to go directly to customers. This will make a to-C enterprise spend more money and take on a lot of pressure in customer service and related support. The use of intelligent conversation robots to provide new retail service chains can reduce user wait times, improve user experience, and greatly reduce the company's customer service costs. The most widely used field of intelligent dialogue robots is online customer service, followed by intelligent question and answer, such as intelligent medical treatment.

From a technical perspective, the flow of human-machine dialogue is: Speech Recognition (ASR), Text-Based Natural Language Understanding (NLU), Dialogue Management by Understanding Intentions or Entities (DM), Natural Language Generation (NLG), and Text To Speech (TTS).

Companies can choose Rasa Stack as the basis for building intelligent conversational bots, an open-source, machine-learning-based robot designed for developers and companies with high intelligence. Because the software development of dialogue robot is very difficult, the understanding of natural language requires the cooperation of many components, and the advantages of Rasa Stack are complete data control, self-expansion, custom model and complete self-driving, and it is backed by Germany's Rasa Technologies GmbH, which has certain development guarantee.

Rasa NLU can extract user intent and related entities, which is equivalent to transforming users 'strange, unstructured, and varied data into structured data. Rasa NLU is characterized by pipeline-based working mode, strong scalability, and support for multiple languages, such as English, German, Chinese, Japanese, etc. RASA NLU also has built-in algorithms and configurations, such as MITIE, CRF, Embedding, etc.

RASA Core is a dialogue management system, as shown in the figure below, where each arrow represents the flow of data.

Rasa Core features data-driven, scalable, support for multiple Policy collaborations, built-in algorithms and configurations, and support for Interactive learning.

In fact, reinforcement learning is not only a framework, it also provides algorithms and configurations, but specific practices and parameter adjustments need to be determined in practice scenarios. In addition, interactive learning can quickly test whether the user gets the correct response and make changes in the wrong case.

"AI Enabling Marketing Digital Transformation" Li Zezhou Pinlan Co-founder and CTO

Currently, offline marketing is shifting from a retailer-centric to a shopper-centric form. Among them, the landing of machine vision has also greatly promoted the transformation of the entire retail industry.

Computer vision is an interdisciplinary field that deals with how computers can gain high-level understanding from digital images or video. Advances in hardware and algorithms have spawned a large number of computer vision applications.

Before deep learning entered the realm of computer vision, computer vision techniques were primarily applied to image processing, feature detection and matching, and motion estimation.

With the development of deep learning networks, it is difficult for traditional neural networks to be simply applied to the field of computer vision. The latitude of the image is very large, and the way of human observation is to observe a certain local information in the image carefully before gradually observing the global information.

The process of machine learning is data acquisition, data preprocessing, model training, model testing and model service. Among them, retailers in the retail industry are more concerned about how SKUs are placed in supermarkets.

Offline retailers have two ways to collect data. The first method is to collect data by using handheld terminals (SFA); the second method is to set up cameras in supermarkets to shoot fixed scenes. Compared with mode 1, mode 2 has certain advantages, such as high-pixel camera can be selected, image quality is high; shooting in fixed scenes has little change; data availability is high; the model can ensure the recognition of fine-grained goods.

In the data preprocessing stage, there are also two ways to process the collected image data information. One way is to improve the image quality, such as adjusting brightness and contrast, deblurring and super-resolution reconstruction of the image.

The second way is to train image enhancement. In the implementation process of AI, before training CNN network, data enhancement is a very important link. In general, the amount of data that can be collected in real scenes is not very large, which is a fatal problem for deep learning. At this time, image enhancement or expansion of image data can be used to increase the amount of data, such as image cropping, image contrast change, image brightness change and image micro-rotation.

In the training phase of commodity detection model, the current frontier algorithm frameworks include Faster-RCNN, RetinaNet and YOLO.

The general commodity detection model can support a large number of actual scene data of various packaging categories, can split large categories for large and small targets and different packaging categories, and can optimize the model structure and enhance the scene adaptability. At the same time, the universal commodity detection model can realize mobile compression and support mobile detection.

The fine-grained product recognition model training needs to collect a large amount of SKU data, establish a product database, and then combine the attention mechanism to train the fine-grained recognition model, and then verify the model effect in real scenes. In practice, Pinlan's fine-grained commodity recognition model has been able to increase the accuracy of commodity recognition in natural scenes to 97%.

After establishing the detection model and recognition model, developers can combine the two to perform intelligent display analysis of offline retail, such as display location inspection, quantity inspection, display specification inspection and display recommendation.

The data-driven new retail era has come. Facing the transformation of retail places and consumption concepts, traditional retail needs to be integrated and reorganized, making full use of e-commerce, big data cloud platform, mobile Internet and artificial intelligence to make online and offline integration possible.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.