The basic recommendation of content-based recommendations is: Users like fantasy novels, this book is fantasy fiction, then users may like the novel Two requirements: (1) Know the user's preferences; (2) know the properties of the item
Content-based recommendations compared to the collaborative filtering approach (personal view): Collaborative filtering uses a large number of users of the group behavior characteristics, two characteristics, (1) to a large number of users, (2) In addition to the user's behavior, no other information; Content-based recommendations require additional information about users and items such as user preferences, item attributes, and so on , but there is no need to store and process large amounts of user data.
Content-based recommendations and knowledge-based recommendations do not have a clear distinction between the two: The former focuses more on extracting item attributes, the latter focusing more on causal relationships.
Content-based recommendations are generally used in articles, news recommendations above. The properties of the article are automatically extracted from the text content. User preferences are extracted from previously evaluated articles by users. General keywords to represent the article and the user.
1. Content representation and similarity
The basic idea of content-based recommendation is to calculate the similarity between the item and the item that the user likes, so as to determine whether the user likes the item and how to get the user's preference. (1) A clear question to the user; (2) The user is asked to rate a series of item, which automatically extracts the feature classic method: Using keywords to express text, TFIDF the weight of keywords, VSM model to calculate the correlation between text problems: How to choose the most representative of the N Words (feature selection). Cons: Inability to use Word context information
2. Search based on content similarity
The collaborative filtering method can be described as "recommending similar items to the user's liking"; Content-based recommendations can be described as "items that suggest something similar to what the user used to like". This section, based on the previous section, discusses how to recommend items to users.
2.1 k Nearest Neighbor method for items not seen by a user, find the user has previously evaluated the most similar to the current item k, according to the user's rating of the K items, to vote to solve the current unknown items score. Billsus2000, personalized Mobile News recommendation: KNN is used to model the short-term interests of users, using probabilistic classification models to model the long-term interests of users. Two models together to give users personalized news recommendations. How to combine short and long interest models? Method 1: Prioritize short-term models, then long-term models, Method 2, use short-term models to find all relevant item, and then use the long-term model for comprehensive sorting
2.2 Relevance Feedback The--rocchio method requires the user to explicitly give feedback to the item during use, which is not practical. In practice, you can consider getting user implicit feedback. Personalized search? How to integrate different correlation systems (PageRank and user and item correlation) is a problem.
3 Other text classification methods
Turn the referral problem into a classification problem, using a classifier to make recommendations, such as: naive Bayesian + Two classification, to predict whether the user likes a particular commodity and SVM, decision tree and other models. Involves feature selection, using Chi-square selection or Fisher discriminant.
4. Summary
Content-based recommendation technology, most methods are derived from the field of information retrieval.
Finish.
"Book Notes" recommendation System (Recommender systems An Introduction) Chapter III recommendations based on content
