【WOT2018】Recommended search venue reveals the secrets and discusses how intelligent search can improve business value

[51CTO.com original article] In 2018, artificial intelligence, as one of the most dazzling technologies for mankind to create the future, is profoundly affecting the global industrial structure, business model, urban form, and human life and work methods. How to enable artificial intelligence technology to empower the industry and let many industries share scientific and technological productivity is an important topic under the wave of digital transformation. As an important annual event to promote pragmatic innovation in artificial intelligence, the WOT2018 Global Artificial Intelligence Technology Summit was grandly held at the Beijing Yuecai JW Marriott Hotel from November 30th to December 1st. 60+ domestic and foreign artificial intelligence elites and more than a thousand industry professionals gathered on site to share artificial intelligence platform tools, algorithm models, voice vision and other technical content, and discuss how artificial intelligence can give new vitality to the industry.

On the morning of December 1st, the Recommended Search Forum was held at Venue A, and three senior experts were invited to attend and give wonderful speeches. After the meeting, 51CTO compiled the experts' speeches into a text, hoping that their understanding of technology and solutions in practice can be used as a reference for everyone.

Zhang Kangqihu Technical Manager of 360 Artificial Intelligence Research Institute

Application of Deep Recommendation System in 360

The main content of Zhang Kang's speech was to share the application of deep learning recommendation systems in various scenarios of 360. He said that the most superficial recommendation system is modeled through functions. If you need to go deeper at the abstract level, you can split the simple mathematical formula into different algorithm modules, such as recall, sorting, and strategy. The last level is to serve the recommendation system online, which is mainly composed of five parts: ETL (data cleaning), Server module, Platform (machine learning training platform), A/B testing, and reporting.

He believes that in real-life deep recommendation systems, 80% of the content in many papers will focus on the description of the model part, and only 20% of the content is left for experiments. In fact, experiments contain rich content, but they are easily overlooked. He gave an example, saying that data analysis and cleaning, feature design and crossover, model comparison, recall strategy, and re-ranking strategy all have a lot of content that needs to be practiced, and the real traditional experiments such as evaluation, A/B, and experimental analysis are actually intertwined with the previous contents.

The application of deep recommendation system in 360 mainly has three aspects: APP recommendation (360 Mobile Assistant), frequently searched words (360 Search & Navigation), and personalized message push (Quick Video). APP recommendation is the first attempt of their entire team in the recommendation field. They finally established a multi-classification model and introduced the image search solution to predict whether the user installation is accurate. Finally, they handed it over to the business department for agent optimization, and the feedback was relatively good. "Frequently searched words" are words that users often search. Zhang Kang introduced that 360 Navigation has hundreds of millions of PVs every day, so they need to perform daily offline updates and perform statistical features based on users' search logs and browsing logs. After personalization + feature optimization, the revenue of this part of the business increased by 5%.

The starting conditions and time conditions for personalized message push are related to the user's personalized real-time. It is not a recommendation at a fixed time and place every day. The content is related to user behavior and requires personalized recommendations, which is of great significance for "pulling activity". They first optimized semantics, vision, and behavioral recall, but later found that the user experience was not good. They also needed to optimize the recall data flow, update the model, adjust the online service, and do a lot of work. In addition, the visual field, semantic model, and sorting must be optimized. In terms of architecture, there must be an offline part to implement downgrade services, log collection, and reporting, and it must also interact with the business side's online system in real time to give the business more active users.

Jianqiang Wang, founder of Stitch Fix, former senior technical director of Twitter

Stitch Fix: A last-ditch fight based on algorithmic recommendations

Wang Jianqiang told the audience that Stitch Fix is ​​a data-driven new clothing retail company, 100% of its sales come from recommendations and algorithms, of which about 35% of sales on Amazon come from recommendations, and 50% of LinkedIn users come from recommendations. Stitch Fix uses a blind box sales model, that is, users do not know what clothes they receive before they receive them. "We will deliver five pieces of clothing to users and deliver them to their homes by express delivery. Users keep the ones they like and return the rest for free." Wang Jianqiang said that the company's trial and error costs are relatively high, and there are also two-way logistics expenses. Based on this, the accuracy of algorithm recommendations is an important indicator of Stitch Fix's corporate survival.

According to him, Stitch Fix has four features to help those working professionals who don't have time to go shopping or are not very good at dressing, namely, fine-grained recommendations, low frequency, self-built inventory, and human-machine coupling. The effect of human-machine coupling is 1+1>2, and a large number of inventory SKUs can be screened and sorted through machine calculations, and patterns can be found from large-scale data, reducing the pressure on manual work.

Wang Jianqiang gave an example of the actual application of the algorithm recommendation model on the spot. When a stylist recommends clothes to a user, he sees the user's relevant data, such as this user is 30 years old, a mother, wears an extra small size, and lives in Minnesota, a state in the Midwestern United States. The system divides the user's style into seven dimensions, each of which is scored from one to four. The stylist will judge the user's style preference based on this score. Finally, based on the body size data, the stylist can recommend clothes to the user with greater confidence. "In the main recommendation model, the model we are using now is a mixed effects model. There are two reasons for choosing a mixed effects model. One reason is that our team mainly has a statistical background. Another reason is that the mixed effects model will have higher technological execution and higher accuracy."

Jiang Qiancheng Senior Algorithm Technology Expert at Meituan

Deep Learning Practice of Meituan O2O Service Search

Jiang Qiancheng said that Meituan covers all areas of users' food, accommodation, travel, shopping and entertainment, with 310 million active buyers per year. Because of the wide variety of services, there is also a problem of how to quickly find the content that users want, which is the value of search on the Meituan platform. It is understood that service search will cover about 40% of transactions, and the number of user products served every day is at the billion level.

Compared with traditional search platforms, Meituan's search platform has many characteristics and challenges of its own. Jiang Qiancheng introduced that, first of all, users have higher personalized needs, secondly, the needs are scenario-based, and the data and retrieval levels are very different in different scenarios. Thirdly, Meituan users have non-standard attributes, and products are not standardized. Even for the same dish, different users like it differently. In addition, it is real-time. The demand for food in the morning at work and at night at home is different.

So how does Meituan solve these challenges? Jiang Qiancheng said that in summary, it is through two paths - finding what users want through intelligent matching technology, and how to enable users to find it more quickly. The essence of these two points is actually recall and sorting.

There are two points to note in intelligent matching technology. One is the matching of user intent, and the other is the matching of surroundings. Jiang Qiancheng gave an example. For example, at Beijing South Railway Station, users have a greater demand for catering, but if it is their first time in Beijing or they are on a business trip occasionally, then they may have a greater demand for accommodation. So even in the same location, Meituan will present different content in different scenarios. "In terms of multiple dimensions, in addition to traditional text matching, we have also added the idea of ​​​​vectorized recall." Jiang Qiancheng explained that after doing a good job of user matching, even if a hundred contents are recommended to users, users will only look at a few at most, and will leave if they are not satisfied. Therefore, it is necessary to find the content that users want most through personalized sorting.

Personalized sorting mainly has three levels: recall layer, model layer, and business layer. The model layer has evolved to the present and has entered the real-time and deep learning stage. Real-time needs to be achieved through feature real-time and online learning. Jiang Qiancheng focused on the framework of Meituan's deep learning, which can better support online support for very large-scale data and models, support for multiple model definitions, and support for streaming model training. "In the end, Meituan's orders have been significantly improved. In various business lines, our current main models are deep learning models and streaming models, which have very positive effects." Jiang Qiancheng concluded.

The above content is compiled by 51CTO reporters based on the speech content of the "Recommended Search" sub-forum of the WOT2018 Global Artificial Intelligence Technology Summit. For more content about WOT, please pay attention to .com.

[51CTO original article, please indicate the original author and source as 51CTO.com when reprinting on partner sites]