Cold start problem of recommended system

The recommendation system needs to predict the user's future behavior and interest according to the user's historical behavior and interest , So a lot of user behavior Data becomes an important part and prerequisite of recommendation system . How to design a personalized recommendation system without a large amount of user data and make users satisfied with the recommendation results so that they are willing to use the recommendation system , It is the problem of cold start . There are three main types of cold start problems , User cold start , Item cold start , System cold start .

User cold start

User cold start mainly solves the problem of how to make personalized recommendation for new users . When new users arrive , We don't have data on his behavior , So it's impossible to predict his interest based on his historical behavior , So I can't make personalized recommendation for him .

Use user registration information

The personalized recommendation process based on registration information is basically as follows ：
(1) Get the user's registration information ;
(2) Classify users according to their registration information ;
(3) Recommend to users the items they like in their category .

Choose the right item to start the user's interest

Another way to solve the user cold start problem is when a new user first accesses the recommended system , Don't show the user the tweet immediately Recommendation results , But to provide users with some items , Let users feedback their interest in these items , Then, according to the feedback from users, I will give you

Item cold start

Item cold start mainly solves the problem of how to recommend new items to users who may be interested in it for personalized recommendation . Item cold start is very important in news websites and other websites with strong timeliness , Because there are new people's items in those websites all the time , And every Items must be able to be displayed to users at the first time , Otherwise, after a period of time , The value of the goods is greatly reduced .

about UserCF Algorithm

about UserCF The algorithm needs to solve the problem of the first driving force , Where the first user finds new items . As long as there is a small part People can find and like new things ,UserCF The algorithm can spread these items to more users . The easiest way to solve the first driving force is to randomly display new items to users , But it's obviously not very personal , Therefore, we can consider using the Content information , Put the new item first to users who have liked other items similar to its content .

about ItemCF Algorithm

about ItemCF Algorithm , Cold starting is a serious problem . because ItemCF The principle of the algorithm is to recommend items similar to the items he likes before ,ItemCF Every once in a while, the algorithm uses user behavior to calculate the item similarity table （ It is usually calculated once a day ）. therefore , When a new item is added , This item will not exist in the item related table in memory , If new items are not displayed to users , Users can't act on it , thus ItemCF Unable to recommend new products . So , We can only use the content information of the item to calculate the item related table , And update related tables frequently .

System cold start

System cold start mainly solves how to build a new website （ No users yet , It doesn't work Household behavior , There's only some information about the items ） Design personalized recommendation system , Thus, when the website is just released, the user body Experience the problem of personalized recommendation service .

Play an expert role

A lot of recommendation systems are built , There is no user behavior data , There is not enough content information to calculate the exact similarity of items . that , In order to let users get a better experience when the recommendation system is established , Many systems use expert tagging .
as everyone knows , It is difficult to calculate the similarity between music . First , Music is multimedia , If we calculate the similarity between songs from the audio analysis , The technical threshold is very high , And it is difficult to calculate satisfactorily . secondly , Just use the album of songs 、 It is difficult to obtain a satisfactory song similarity table from attribute information such as singers , Because a singer 、 An album often has only oneortwo good songs . To solve this problem ,Pandora hire
A group of computer literate musicians were used to carry out a project called music gene . They listened to the songs of tens of thousands of singers , And annotate each dimension of these songs . Final , They used 400 Multiple features Pandora Call these traits genes ）. After marking all the songs , Each song can be expressed as a 400 Dimension vector , Then the similarity of songs can be calculated by the common vector similarity algorithm .
Jinni In the film genetic engineering, semi artificial 、 Semi automatic mode . First , It allows experts to mark films , Every movie has about 50 A gene , These genes come from about 1000 A gene bank . then , After experts mark certain samples ,Jinni Be able to use natural language understanding and machine learning technology , By analyzing the user's comments on the movie and some content attributes of the movie ( Especially new movies ) Make your own mark . All in all ,Jinn The cold start problem of the system is solved by combining expert and machine learning .