It turns out that there are so many tricks in O2O search! Meituan Xian Yunsen reveals how to make services smarter

[51CTO.com original article] The Global Software and Operation Technology Summit hosted by 51CTO was held in Beijing on May 18-19, 2018. The summit focused on 12 core hot topics such as artificial intelligence, big data, the Internet of Things, and blockchain, and brought together 60 front-line experts from home and abroad. It was a high-end technology feast and a platform that top IT technical talents should not miss to learn and expand their network.

At the "Artificial Intelligence Technology Exploration" forum in venue B on the morning of May 18, Xian Yunsen, head of Meituan's search algorithm, gave a keynote speech on "Intelligence of O2O Service Search" and demonstrated some research results and insights on the basic algorithms of Meituan's platform search technology center.

[[230895]]

The value of O2O search: connecting people and services

Xian Yunsen first introduced the overall characteristics of O2O (Online to Offline) search and some differences between it and web search or shopping search. He introduced that Meituan is a life service e-commerce platform with dozens of business lines, and there are very large differences between each business. He listed three businesses for comparison. The first is foot massage. Foot massage is a very light decision-making business. Users pay more attention to the location. The proportion of consumption within one kilometer is very high, and the experience is not important. In comparison, home decoration is different, because the decoration will be used for many years, and users are less sensitive to distance and discounts, and pay more attention to quality. Marriage focuses more on quality and user experience. These different search needs bring great challenges to the search business.

What is the mission of search? Xian Yunsen believes that search is to connect people and services, and connect users and businesses. If we only use two words to describe it, it is "efficiency". "There are two meanings here, one is transaction efficiency, and the other is decision-making efficiency." He believes that search has two values ​​for merchants: one is user reach, which means getting more users to the store through search, and the other is conversion efficiency, which means getting users to consume in the store and optimizing while placing orders.

Xian Yunsen also compared O2O with web search and shopping e-commerce from four levels.

The first difference is mobility. Meituan’s mobile terminal accounts for more than 94%, Taobao’s mobile ratio has also reached 90%, and Baidu’s mobile ratio is slightly weaker. The second difference is personalization. Baidu is lower in this comparison because Baidu’s philosophy is that "information is equal and does not need personalization." The personalized options of Meituan and Taobao are very high. The third difference is scenario-based. For example, booking hotels and tickets for scenic spots are often reached in different places, and the impact of geographical space is very small. The fourth point is real-time. Xian Yunsen believes that Taobao’s real-time is less related to scenarios, but the real-time nature is enhanced in specific scenarios such as Double Eleven and anti-cheating.

How to make O2O search intelligent?

After understanding the characteristics and basic features of O2O search, how can we realize the intelligence of O2O search? Xian Yunsen believes that there are two goals of intelligence: one is to reach precise needs with one click, and the other is to improve decision-making efficiency by optimizing experience. There are also two key paths: accurate understanding of user intentions and intelligent sorting algorithms.

In the search technology panorama listed by Xian Yunsen, from the basic layer, strategy layer to the business layer, he sorted out the search strategies and technologies corresponding to general business and vertical business respectively.

At the event, Xian Yunsen also emphasized the role of intent understanding, that is, how to understand the user's intention. He introduced that the user intention on Meituan can be broken down into several key elements, such as location and time. The location can be solved by searching the surrounding area. The needs of users at different times are also different. For example, at noon on weekdays, users may only have a demand for fast food takeout, but on holidays, they may search for feasts. In addition, weather, rainy, snowy, foggy and hazy weather will affect users' consumption behavior.

When talking about intent and recall, Xian Yunsen said that Meituan is a structured search. Unlike web pages, each store on the Meituan page has several fixed fields, and all are standardized remote manual orders. When Meituan builds an intent system based on intent, there are four very important points. First, business identification must use multi-intent identification. Second, the consumption location needs to be identified as local or remote. Third, the consumption time must be determined whether it is immediate consumption or delayed consumption, depending on whether the user has a queuing demand or a reservation demand. Another point is component analysis. If the search system can clearly identify its components, it can recall more accurately and have a better chance of accurately meeting user needs.

For example, when the user enters the term Beijing South Railway Station, the demand for hotels is overwhelming. If the user enters the term Sanlitun, the demand for bars will be very large. When searching for Wangjing and Wudaokou, there will be a certain demand for Korean food. "This is how we can judge the distribution of different intentions based on some real behaviors of users."

Deep learning brings improved results

At the end of the speech, Xian Yunsen also shared Meituan's practice in real-time. They first made features real-time, and did some special engineering on users' real-time behavior, real-time inventory, real-time intentions, and real-time conversions. Their model is actually based on historical data to discover patterns and predict the future. If only the data has changed but the pattern has not changed, then feature prediction is enough. Later, they began to try real-time models and went through three rounds of iterations. The first was the basic model GBDT+LR (Gradient boosting decision/regression tree, a nonlinear model that is widely used in LTR), which is also Facebook's more classic structure. On this basis, they made some improvements and added the features of sbas directly to the LR end, which further improved the effect.

When talking about deep learning, which is very popular nowadays, Xian Yunsen said that deep learning is of great help to Meituan. In the past, from the perspective of business characteristics, there were many cross-search factors such as user scenarios, user status, user queries, etc., and it was not enough to rely solely on manual power and tightening force. Later, they tried to use deep learning, refer to the DeepFM model, add some feature combinations to wide, and change the FM part to grouped product, which achieved a very positive effect in the basic network. In addition, since Meituan is a service provider focusing on transactions, relevance is also very important. Deep learning is more suitable for multi-task learning, so they tried two structures, one is shared wide and the other is shared deep. In the end, they chose shared wide because the cost of not sharing wide would be greater.

The above content is compiled by 51CTO reporter based on Xian Yunsen’s speech at the WOT2018 Global Software and Operation Technology Summit. For more information about WOT, please visit .com.

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