1, user behavior data
in the site simplest form of user behavior data is a log of user behavior refers to various operations performed by the user in the system, such as the user's browser in the electricity supplier website, click, search, buy, collections and other behavior. We can speculate by analyzing the data which the user's favorite commodity, so they recommended a more preferred commodity for the user.
2, classification of user behavior
According to the explicit feedback, the user behavior in a personalized recommendation system generally two types: ① explicit feedback behavior: it clear that the user likes and dislikes of goods behavior, such as assessment, collection and so on; ② implicit feedback behavior: Can not made it clear that the user likes and dislikes behavior, such as through the pages and so on.
According to the direction of the feedback, the user behavior data can be divided into: ① positive feedback: For a commodity users tend to like it; ② negative feedback: For a commodity, users tend to dislike it.
For explicit feedback behavior we can easily distinguish a user behavior is positive feedback or negative feedback, and in the implicit feedback behavior, it is relatively difficult to determine.
3, recommendation algorithm
In the recommendation system, the most basic and widely used algorithms are user collaborative filtering algorithm (UserCF) based and collaborative filtering algorithm items (ItemCF) based . User-based collaborative filtering algorithm (UserCF) The basic idea is: when recommended A user, you can find other users and he has similar interests, then these users like and A user does not browse the items recommended to A, that is, a recommendation for the user and he has similar interests of users like commodities. UserCF recommendations focus on results reflect the user's interest and similar hot small groups, and ItemCF recommended to maintain user focus on the results of historical interest. In other words, UserCF recommended more social, reflecting the popularity of small items in the user's interest groups, and ItemCF more personalized recommendations, reflecting the user's own interests. So, news sites and more UserCF algorithm, multi-site electricity supplier recommended by ItemCF algorithm.
But with the increase of users of the system, user interest similarity matrix computing will become increasingly difficult, the computational time complexity and the number of users increase growth and complexity of space similar to the square. And user-based collaborative filtering recommendation is difficult to explain the results. Accordingly, another method is proposed Amazon - collaborative filtering algorithm article (ItemCF) based.
Collaborative filtering algorithm based on item (ItemCF) The basic idea is: the user is recommended to those items and similar items before they like it. For example, the algorithm because you bought "Java from entry to the master," and recommend you for "Java Concurrency in combat." However, the degree of similarity between the items calculated based on the content attribute collaborative filtering algorithm using the article is not an article, is the similarity between two items is calculated by analyzing the user behavior data. The algorithm considered, item A and item B has a great similarity is because I like the article A user's most liked items B.
ItemCF process:
① calculating the similarity between the goods
N (i) on behalf of the user number i like product, whereby we can see, W is ij of the representative like the user i while the items the user likes ratio j of goods.
After obtaining the degree of similarity between the goods, ItemCF user interest in a product calculated by the following equation:
Where N (i) is a collection of items the user likes, S (j, K), and is most similar commodity j K sets of articles, W ij of a similarity items i and j, X- UI user u commodity i love the extent of.