Collaborative filtering is a typical way to use collective wisdom. First, think of a simple scenario. If you want to see a movie now, but don't know which one to watch, what will you do? Most people will ask their friends to see what good movies are recommended recently, but we generally prefer to be recommended by friends with similar tastes. This is the core idea of collaborative filtering.

The following figure takes the user's recommendation of movies as an example. Table 1 is an example of a simple matrix of user rating data for movies, where each row represents a user and each column represents a movie. The elements in the matrix represent the user's rating of the movies watched, and the rating value is generally an integer from 1 to 5. The greater the score, the more users like this movie.

Using collaborative filtering recommendation algorithm for the data in table 1, the system found that users Alice, Bob and Chris have similar interests and hobbies, because they have the same scores for the last three movies, so the system will recommend Snow White to Chris, because users with similar interests and hobbies Alice and Bob have higher scores for this movie. In Table 2, there is no rating information of Amy, a new user. According to the collaborative filtering recommendation algorithm, users with similar interests and hobbies can't be found according to the scoring information, so the system can't recommend movies for this user. Similarly, the new movie Shrek can't be perceived by the system because of the lack of rating information, so it can't be recommended. This is the problem of new users and new projects in collaborative filtering recommendation algorithm, that is, the cold start problem.

"Cold start" includes two levels:

1. The user's cold start, because the new user didn't leave behavior data on the product, naturally he/she can't know his/her preferences, so he/she can make reliable recommendations. At this time, it is generally necessary to rely on the user's background information, or guide the user to choose, or temporarily replace personalized recommendation with popular startup. The online recommendation system can update the recommendation list immediately after the user generates behavior data.

2. Cold start of items is mainly suitable for item-based scenarios, that is, pushing similar items to an item. Because there is no user behavior data of new products, there is naturally no way to recommend them through collaborative filtering. At this time, it is generally solved by the correlation of item attributes.

In real life, for e-commerce recommendation system, a large number of new users visit the system every day, and a considerable number of new products join the system every day. On the one hand, if the recommendation system can recommend favorite products for new users, it will win the trust of more users, increase the number of customers for merchants, improve the loyalty of users to the system, and provide users with high-quality personalized services at any time; On the other hand, if we can recommend new products in time, we can increase the sales of products, win greater economic benefits for businesses and promote the healthy development of e-commerce.

At present, some solutions to the cold start problem are put forward, such as random recommendation method, average method, pattern method, information entropy method, similarity measurement improvement method, content information combination method and so on.

This article is taken from/s/blog _ 7F9bc990102W81l.html.