Item-based and user-based selection basis in collaborative filtering _r



This article summarizes the good.

Collaborative filtering is a well-known recommended algorithm. In general, collaborative filtering can be divided into the following two categories: neighborhood-based: Calculate similar item or user to recommend model-based: Direct training Model Prediction Rating

In the neighborhoold-based algorithm, it is subdivided into user-based CF (collaborative filtering) and item-based cf. Appropriate choice to use userd-based CF, when item-based CF is more applicable will be a need to weigh the problem. In general, you can choose from the following criteria: accuracy: Generally speaking, the recommendation of a few trusted neighbors is more accurate than a lot of neighbors without much distinction, so we usually choose a small number of factors (item or user) as the based algorithm. For example, there are many kinds of products in Amazon, but far from registered users, so the scene using item-based CF is more appropriate, in turn, in the Baidu keyword recommendation system, commercial customer (user) level is about 100W, while the recommended keyword (item) is 1 billion magnitude , it would be a wiser choice to use user-based at this time. Efficiency Stability: In general, the tendency to use the variable frequency and less variation of factors as based factors, such as the item change less, then choose item-based, otherwise choose user-based Justifablity (Persuasive): Recommendation system, the more white box, the more users understand the more persuasive. So from this point of view, item-based CF will be more persuasive, for example, ' because you're browsing the Samsung Galaxy, so you're recommending HTC One ' for a better reason than ' Similar to your users also like xxx ' more convincing, because the recommendation system is not disclosed which users and I detailed, how to prove similar to me, and this statement seems more ambiguous. Serendipity: Diversity is one of the big advantages of user-based, and users like yourself can always find something new that they haven't found yet. If you pursue diversity, userd-based will be a good choice.

Of course the above principles are not absolute, and in the real industry recommendation system, the two methods are generally mixed use. For example, Baidu keyword recommendation system, will be used respectively item-based and user-based method to find the Recommended keyword candidate, and then unified use model for follow-up ranking.

References: RSs Handbook evaluating Collaborative Filtering recommender Systems, Jonathan l.herlocker