Talk about beer and diapers.

"Sir, I noticed you bought a beer, would you like some more diapers?" "If the supermarket cashier asks you this, will you want to K-man?" Even if you know the story about beer and diapers.

In fact, the original story is that many men were found to buy diapers in Friday, so the supermarket put beer beside the diapers. So I just deliberately confused two details: diapers and beer are "one-way", the recommended way is "implicit" (put together) rather than "explicit" (oral recommendation). More importantly, it is about the behavior of a specific target user group (pecked men who are married and have children).

Surprisingly, I've seen most of the "beer and diaper stories" deliberately ignoring these details and turning to big talk about collaborative filtering or data mining. I'm afraid that's why so many of the websites ' recommender systems are doing pretty bad things right now. Recommendation system, essentially a product, rather than what the technical framework, return to the fundamental principles of product design, it is possible to do it "useful", and then "easy to use."

I am in fact also to recommend the system's technology is not good, might as well put aside those technical algorithms, talk about some I think very important design principles:

One, focusing on "follow-up expectations". For example, suppose I'm browsing the ipad product introduction, what should I recommend next? Imagine several answers: A:iphone,ipod;b:ipad leather sets and other accessories; C: The ipad quotes from different channels; D: Other Shanzhai products that mimic the ipad. If a classmate want to buy a fashion electronic products to his girlfriend, B classmate is a fan of the ipad, C students are still hesitant for the price of the ipad, ding students just want an electronic reader, the above four answers are exactly enough to allow them to continue to look at a number of products. Therefore, the most critical factor for the success of the recommendation system is not the technical problems such as algorithmic implementation, but how to discern the user's current behavior and subsequent expectation.

It is worth mentioning that, in addition to speculating about the user's own expectations, creating a new expectation for the user is also a recommendation system often done. For example, the website selling books often tells me that these two books can be bought. But this is recommended for users he did not expect things, the success rate will be much lower, on the contrary may disrupt the user's original behavior path. So be careful not to be smart-aleck.

Two, remember the 80-20 rules. To use a set of algorithms to solve the different needs of millions of users is impossible, which requires us to make some trade-offs. I have a bypass principle: take care of most of the most common needs of most people in the simplest way. My inspiration for this principle comes from a desktop search software called modifiable. Whether it's Microsoft or Google's Desktop search software, it always takes a long time (a few hours) to build the index at first. But this modifiable only in 1 minutes to build the entire hard drive file index, very magical! Later found that it can only search the file name, but not the content of the files, and the latter is the other desktop search for long time indexing reasons. The problem is, in fact, most of the time we look for files are looking for file names rather than content, modifiable is the core of the problem, with 20% of the energy to solve the 80% problem, then, the remaining 20% of the problem directly abandoned!

It would be appropriate to use this 80-20 rule in the recommendation system. All you need to do is focus on the most frequently occurring situations and use the appropriate algorithms to find the right recommendations, and the rest of the results are guaranteed to be "relatively passable." Of course, this is done on the premise that you are confident of the user's follow-up expectations.

Third, the resolution of information anxiety. Suppose "Busan Cuisine" Restaurant Introduction page, the relevant restaurant listed "Hong Kong-Li Tea Restaurant", do you think they related? So, "Korean Palace barbecue"? Do you think the latter one is clearly a lot more reliable? The reason is that the latter one you can see it with "Busan cuisine" is a Korean style barbecue shop. But if you are a person in the people's Square about friends to eat, "Busan" and "Port Li" are just right at Renmin Square, and "Han Palace" In fact, is very far from Renmin Square Xujiahui, it goes without saying that nature "Hong Kong Li" should be your next to browse the restaurant.

It's important to tell the user why. Otherwise the user will be confused and ignore what you have carefully recommended for him. A good example is the NetEase news, in its related news list, will it match the label displayed. such as "News related to Foxconn".

Always remember business goals. The ultimate goal of the recommendation system is to achieve business objectives: increase the subsequent conversion rate? Increase the total number of users? or increase turnover? Other product designs may sometimes need to sacrifice the immediate business interests to meet user goals. However, the business objectives of the recommendation system and the user's goals are not divided, but often more unified, so at this time should be particularly reminded of the business goals, all the time to work in that direction. Of course, some very basic user experience or to ensure that, like in the referral system in advertising this kind of thing or do not do, hehe.

Source Address: http://www.mikkolee.com/291