The common concern of hundreds of thousands of Internet practitioners

By Jason and Link, reprinted with permission for Morning Reading

From 0 to 1 (ID: Aboutlink)

Editor: Verna

In July 2015, Wang Xing proposed that the Internet has entered the second half. What is the second half of O2O?

It is not difficult to guess how big companies’ business layout is in the second half of the O2O industry. On the one hand, the O2O industry is mainly engaged in two aspects: the integration of queue, order, bill and other services; on the other hand, the O2O industry is engaged in precise shopping guidance through community and personalization to improve user stickiness.

The author has worked in Meituan-Dianping and Ant for personalized recommendation. Here I would like to talk about the pits I have stepped in the process of recommendation optimization and some business considerations.

There are a lot of contents about how to build recommendation system from 0 to 1, so I will not repeat them here. The content of this paper mainly focuses on the optimization practice of recommendation from 1 to N in O2O situation.

1. Correlation recommendation is better than causation recommendation

Let’s first explain what these two indicators mean:

Relevance recommendation refers to that the user has a certain demand and recommends the corresponding merchant to the user by capturing the demand. Correlation recommendations are further divided into two categories:

  1. The existing needs can be met through more convenient paths, such as recognizing that the user is already at a certain business, recommending the system to put the business at the top, and displaying the available offers, shortening the user out of the large path.

  2. Pick up the interrupted demands of users. For example, the user searched “hot pot” recently, but no transaction was formed, so the recommendation system recommends the merchants of hot pot to the user.

Causal recommendation refers to the fact that users do not have clear needs, and recommendation allows users to create new needs out of nothing.

A controversy in academic circles is “relevant recommendation” is not recommended, because he is satisfy existing demand, no innovation new requirements in a strict sense, my answer is yes, we reverse to push, if you don’t make relevant recommendations, user despite existing demand, but is likely to be lost, or experience is not so smooth. So I don’t think we need to worry about that, we just need to worry about whether the user thinks the recommendation is right or not and how good it is to use it.

Relevance recommendation practice

1) Identify the merchant where the user is

If the user is connected to wifi, it is easy to identify the merchant that the user is in, and then put that merchant on top, and expose the merchant’s discounts and other information. Among all the recommended strategies, the coverage of this strategy is small, but the conversion rate is the highest.

Another way is to add a function called “I want to go” to the merchant details page. After the user clicks this button, the merchant can directly recommend it on the home page within a period of time, shortening the user’s viewing path.

2) Identify users’ intention to visit the store

Through a combination of various behaviors, users can guess which businesses to go to. Imagine if you offer a friend to dinner, what would you have selected merchants, general path see interested merchants are first search, then click go in to see details, including check location, recommend dishes, if you click on the merchant telephone, collect the merchant, or share the merchants to WeChat, then you will go to this store probability is higher, Finally, when combined with your location, you have a very high probability of guessing which store you are going to.

3) Real-time behavior of users

Here to tell two concepts, the user’s long-term preference and real-time preference, we often say the user portrait generally refers to the user’s long-term preference, for example, the author’s long-term preference is Japanese material, but maybe because I like a girl recently, but the girl likes hot pot, so my short-term preference is hot pot. In addition, practice has proved that users’ preferences will decay over time. Real-time preference data can better represent users’ current needs, and the recommendation effect is more accurate.

Real-time preference typically they can navigate by the user’s search, screening or users browse to infer, if the user does not have trading closed loop formation, such as buying a bulk, pay, numeral, etc, we can assume that his demand is not met, may be didn’t find the right store, also demand may be interrupted by other things, At this time, we will recommend users real-time preferences of the merchant to the user conversion rate is often better.

Causality recommended practice

1) Collaborative filtering algorithm

The essence of collaborative filtering based on merchants and collaborative filtering based on users is what Chinese people often say: “Birds of a feather flock together.” The former is to recommend similar merchants to you based on the merchants you like; the latter is to find people with similar tastes to you and then recommend the merchants they like. There are many articles about the differences between the two, which will not be repeated here. The only point to emphasize is that although the two recommendation methods are similar, only about 50% of the recommendation results are the same. Therefore, it is not necessary to choose one of the two strategies in practice, and the two strategies can be combined to achieve better results.

2) Users’ long-term preferences

The user preference here refers to the user’s long-term preference and non-real-time preference. After identifying the user’s long-term preference, the corresponding merchant will be recommended for a period of time.

According to the practice results, relevance recommendation > causality recommendation > supplementary recommendation (nearby popular, whole-course popular), and relevance recommendation is more likely to make users have an impression of “accurate recommendation”. Therefore, relevance recommendation can be given priority in the construction of recommendation system.

2. All love is familiar + accident

At the beginning of my recommendation, I always had a question, that is, what is a good recommendation? I consulted many people at that time, and they were divided into two groups. Another school thinks that a good recommendation is to recommend the shops I am familiar with, because the shops we usually go to are relatively fixed, and I am not interested in recommending the shops I am not familiar with.

Both of these opinions have their own disadvantages. If users do not understand the new content, it is bound to cause poor data effect. However, if users are always familiar with things, it will form Matthew effect, which will become narrower and narrower.

Later, I read an article about a music application software. At the beginning, the algorithm was designed to only recommend music that users had never heard before. However, during the internal test, a bug appeared in the program. In addition to recommending new music to users, it also incorrectly recommended some music that users had already listened to or were even familiar with. This test for a period of time, the effect is good. The programmer discovered the bug and immediately fixed it, making the program recommend only new music. It turns out that the modified algorithm is less popular than the original one. So people don’t like things that are completely unfamiliar, and they always want to find something familiar from something new.

Reading this article, I suddenly understand that good recommendation is to find a balance between novelty and familiarity. How to find this balance, there are several ways:

Starting from the people

If the repeat reading is very low, it indicates that users prefer novel merchants and increase the proportion of new merchants at this time, and vice versa.

Start from the scene

Under normal circumstances, people tend to go to the shops near and familiar with a little on weekdays, and people tend to explore some novel shops on weekends.

3. Recommendation reason is a bridge to improve user trust

Merchants recommended to users, especially those unfamiliar to users, can reduce users’ understanding cost and improve their sense of trust. Click data with recommendation reasons may not be improved (the author passed AB test and the data only increased by 1%), but recommendation reasons can improve users’ browsing experience and understand the recommendation logic. Here are a few cases of some reasons for recommendation:

  1. Friend recommendation: Recommended by x friends such as Jason, it is better to expose the names of friends with high intimacy

  2. Celebrity recommendation: Opened by singer Xue Zhiqian

  3. Media recommendation: Recommended by Star Fashion channel

  4. Making the list: Food & Wine magazine’s best 50 of 2016

  5. People with a group of points: with your taste similar to the people like, modern housewives like

  6. User behavior: Merchants you browsed 7 days ago

4. The highest level of recommendation is to surprise users

What was the surprise? Wiktionary’s definition is:

An unsought, unintended, and/or unexpected, but fortunate, discovery and/or learning experience that happens by accident.

Unsought, unanticipated, lucky and serendipity.

One is music. Some people may have heard a favorite song for a long time, but they didn’t know the name of the song. After hearing it on netease Cloud Music, they would comment that I had found the song for a long time. Netease Cloud Music specially mined such comments, and then recommended these music to users, users will naturally have a great sense of surprise.

The other one is Kindle. One day, the boss told us that Kindle was so amazing that his wife was reading a book and just as she recommended it to him, he found that his Kindle also recommended it to him. We guessed that Kindle is to guess their family relationship by connecting them to the same wifi at night. Then recommend one person’s favorite book to another.

So O2O recommendation how to create a sense of surprise? Here are some starting points to consider:

  1. An old friend in a strange place: when a user is in a strange environment, he can recommend the shop his friends have been to, which is especially applicable in a foreign environment. Just imagine that when you travel to Seoul, Korea, and find a nearby shop your friends have commented on, will there be a kind of intimacy of “an old friend in a strange place”? Of course, stepping back and directly recommending the nearby food list is also applicable, and will produce a certain sense of surprise.

  2. The moon is the Hometown of Ming: The user’s identity labels are skillfully used, such as hometown information. Jason is from Luoyang, where hu spicy soup is our special food. At this time, recommend nearby hu spicy soup merchants to me and indicate the reason for recommendation, or recommend luoyang people’s favorite restaurant collection to me, I will be very impressed.

  3. Love what he loves: It is possible to judge the relationship between the user’s family/friends through the connection of wifi information, etc. At this time, recommending the store that the family/friend likes to another person will make the user surprised, especially when two people go to the same store.

5. Rich recommendation dimensions for all-round hit

Guess what you like personalized recommendation is universal recommendation. Most of the recommendation logic is black box, which is monotonous and not interesting enough for users. In terms of recommendation dimensions, X treasure has done the best. There are 5 kinds of recommendation dimensions, which are complementary to each other in user stratification and content form, and have achieved good results.

Segmentation by user

There are generally two ways of user segmentation: 1) user consumption ability, such as the user group with medium and high customer unit price and the user group with low customer unit price; 2) User stratification, such as student party, hot mom, car owners, etc., can obtain recognition through identity tags, and at the same time meet the needs of users with long tail

Organize by content

There are three types of content organization. 1) Directly promote the store or discount (comment you like), 2) through UGC short comments (some treasure has good goods), and now very popular PGC articles to recommend (xiaohongshu, etc.). Among them, PGC is generally more professional and more suitable for middle and high-end people.

Read More

  • Operation, sometimes is a word

  • Five operating lessons from the resurgence of Sina Weibo

  • All my user experience in Tencent is here

  • Why is the future full stack

  • You misunderstand operations, so you always say, “That operation doesn’t work.”

  • Understand user operation system: user stratification and clustering

  • The open rate of articles is less than 5%, and the number of fans is negative. Why is wechat operation becoming more and more difficult?

Email: [email protected]

This article is published by the author authorized by morning Reading class. Please contact the author for reprinting.

There is no shortage of job opportunities for good people, only good opportunities for themselves. But they often don’t have the energy to sift through a sea of opportunities to find the best fit.

100Offer will strictly screen talents and enterprises on the platform, so that “the best talents” and “the best companies” meet.

Scan the QR code below to register 100Offer and talk about your expectations for your next job. Within a week, you’ll receive 5-10 great opportunities to meet your needs!