Blued is a “different firework” for vertical social products thanks to AI

This is a famous gay love poem of the Spring and Autumn Period. Since ancient times, there have been numerous works depicting same-sex love, but due to the limitations of traditional moral concepts, same-sex love has long been unknown to the public in China. The recent handling of gay content and information on the Internet seems to give the impression that things are still tough for the community. The success of Blued, a social app for gay men, has given this group a degree of belonging and demonstrated the huge potential of the market. Few people know that its success is inextricably linked to advances in artificial intelligence.

It is my great honor to invite Blued data scientist Yingjie Wang to share the topic of “AI Image Technology Application in Social Networks” at ArchSummit Global Architect Technology Summit held in Shenzhen on July 6th. Taking this opportunity, AI Front interviewed Wang Yingjie to make a detailed interpretation of the application and effect of AI in Blued.

(The following is the interview)

Compared with some European and American countries, China seems to be less tolerant of homosexuality, but this has not stopped the survival and development of this group, as well as the growth of consumption power. Blued was born in this environment.

Like other Internet companies, Blued has begun to improve the experience by incorporating the hottest technology of the moment, artificial intelligence, into the daily operations of its products to cope with the growing influx of new users.

Faced with a huge number of users and social network data information, how to find friends around each person and match them according to users’ interests has become a great challenge for Blued algorithm engineers. The importance of images, videos and dynamic pictures for social networking sites is self-evident. A large part of the work of algorithm engineers is to deal with data related to visual information, and AI has become the best tool for them to solve problems.

According to Wang Yingjie, a data scientist in Blued’s AI algorithm department, Blued has widely adopted AI technology in the platform’s image social business.

They cut into the market from stranger social interaction for LGBT people, and gradually turned to interest social interaction and pan-entertainment platform, expanding many application scenarios. In these scenarios, users can publish face profile pictures, photo albums, picture dynamics, small videos and live broadcasts on Blued. In view of the importance of visual information in social products, users are expected to browse photos with high appearance level under the interest tag, get to know each other quickly with small videos, and watch the live broadcast of recommended anchors. Blued’s social, content and commercialization modules have been applied in depth using AI image technology.

In social products, face detection technology is used to screen the profile photos containing faces, and similarity analysis is performed on the obtained facial features. The body image is classified as fat and thin, and some body and clothing labels are extracted through the detection model, which are very important features for the construction of social recommendation product model. Use the image classification technology to eliminate the small video without human; And use image detection combined with image classification technology to extract feature values recommended by anchors.

In terms of cash business, Blued has launched a commercial advertising module based on its content feed stream recommendation products based on images and short videos. The application of image algorithm in avatar authentication and privacy protection is also an important part of realization business such as membership and value-added services.

Specific to AI algorithms of image technology solutions and applications, the internal mechanism of ying-jie wang explains, Blued AI image technology scheme based on product requirements, first of all, apart a few core image task, select suitable network model, such as models for face detection, face recognition model, the image tag detection model, image classification model, etc. After that, a large number of pictures produced on the platform were used for training and fine-tuning, and some basic models were constantly iterated. Finally, these models are combined in different business scenarios, and the output threshold parameters of the model are adjusted at any time according to the test results during the use process. In terms of algorithm mechanism, model iteration, data accumulation and parameter adjustment form three parallel evolution processes. At the same time, the accumulation of data promotes the iteration of the model. After the iteration of the model, the parameters are constantly optimized and adjusted. After the optimization and adjustment of parameters, the data accumulation with better quality is obtained, thus promoting the evolution of () as a whole.

Through this set of technical solutions that have been running inside Blued for half a year, Blued has solved problems and effects that could not be solved by manual audit, manual operation and product rules before. For example, in some product modules with AI technology, UV growth of more than 30%, per capita PV growth of more than 60%, the recommendation success rate is more than 2 times higher than manual selection. At present, Blued’s algorithm model is updated iteratively every month, but there is still a lot of work to be done to improve the algorithm and optimize the target in terms of the cooperation with content production and the social transformation tendency of content consumption.

To see how the product works, AI Front put it through a test run. After registering a Blued account, the system will make recommendations through the interest TAB selected by the user. So how does Blued’s recommendation ranking mechanism work?

AI Frontier understands that Blued data platform will collect basic information of users’ registration and generate interest tags based on users’ content browsing behavior on the platform. Blued will further explore users’ social relationship chain and import these data into the recommendation system. In addition to the recommendation algorithm, Blued also considers user-defined filtering and filtering criteria for sorting, but mainly based on login time and distance.

Blued is not an ordinary website. It is used by a special group of users, and therefore has some characteristics that are different from ordinary websites, and thus presents engineers with “unusual” challenges. Blued’s AI journey hasn’t always been smooth, and in many cases, engineers have faced an overwhelming number of challenges.

Wang Yingjie admitted to AI front that the biggest technical bottleneck of Blued at present is the parallel computing of large-scale data in cloud and the computing efficiency of mobile terminal model. The difficulty of the former lies in the fact that the model computing platform and data storage platform have not been connected yet, which is already being solved by cloud computing services. The difficulty of the latter is that the current scheme has not reached a good balance between efficiency and performance, because the mobile terminal has high requirements on computing power and power consumption. But Wang believes that with the rapid development of mobile technology, this bottleneck will soon be overcome.

Blued users also have different characteristics, including greater subdivision of interest tags, higher difficulty in identifying the authenticity of user data, more uneven distribution of user feedback behavior, and more frequent visits by users. All these put forward more challenges to the data and computing power of the algorithm.

And these problems are not without solutions. Blued solves the challenge of data problems by extracting more features and experimenting with various clustering and classification algorithms, especially models that are insensitive to missing data and models that do not rely on user feedback behavior. In terms of calculation force, the calculation pressure is distributed among off-line calculation, near line calculation and online calculation, and the calculation frequency and calculation amount of each part are adjusted at any time according to the data.

In addition, social networking sites are often a “disaster zone” for pornography. As an app that mainly serves LGBT people, Blued also undertakes the task of informing users about the prevention and treatment of AIDS and other diseases. Blued also faces such challenges, specifically reflected in pornographic images, text, vulgar content recognition and other tasks.

And Blued in community management by artificial identification standard of strict audit team, in the process of model training and reasoning considering different classification test category in terms of accuracy and recall rate of different requirements, such as pornography need higher accuracy of detection, sexy content detection need higher recall rate, This in turn improves the review efficiency of the human review team. Blued told AI Front that they face a bigger challenge in identifying vulgar content, which is reflected in 1. Judgment criteria will change greatly over time, and change quickly, need to constantly increase or decrease the need to detect the categories; 2. It is difficult to accurately mark samples, and it is difficult to guarantee the accuracy and recall rate of models. At present, Blued is trying to solve this problem by improving the dynamic update process of the model, increasing the strength of manual audit, and increasing the entrance of user report feedback.

The layout of products and services using AI technology has already been laid out, and more technology will be explored in the future.

Blued’s technical planning is to give priority to AI, emphasize the personalized operation of subdivided groups, and digitize and model interest and social knowledge. Based on the social needs of different types of segmentation groups, reasonable product scenarios are designed, appropriate features are found, matching models are selected, and how to select positive and negative samples and refined optimization objective functions are designed. In this process, new product ideas become possible, and experience and knowledge of product and operation are digitized during model training.

In the future, AI technology will be reflected more and more in Blued products, not only in social areas, but also in new commercial opportunities, such as the combination of new social and new e-commerce.

Wang Yingjie, Blued data scientist, is currently working in the AI Algorithm Department of Blued (Beijing Blue City Brothers Information Technology Co., LTD.), responsible for the work related to image and recommendation, including the AI technical solutions and implementation of social, content, live broadcast and risk control, etc. He graduated from Beijing University of Posts and Telecommunications with a PhD degree in 2007. He has a number of domestic and foreign patents and rich experience in image deep learning and image processing technology.