Overview
Recommendation System is a facility of predicting user responses to options. Examples:
- Offering news articles to on-line newspaper readers, based on interest(topics) prediction.
- Online retailer suggestions, based on history purchase/search. Note: physical stores V.S. online institutions, not possible to tailor the store to individual customer, choice governed by overall popularity(sale performance).
Content-based Recommendations
Similarity of Items is determined by measuring the similarity in their properties. Why called content based? Goal is to create both an item profile consisting of feature-value pairs and user profile summarizing the preference of a user.
Item profile
Collection of records representing important characteristics of that item. E.g. stars of a movie, released year, genre… these are features that are tend to be readily available, how about documents? One possible solution is to choose words whose TF-IDF higher than a given threshold to be the feature set.
User profile
Item profile is vectors describing items, user profile is vectors with the same components that describe the user’s preferences–need some aggregation.(naturally take the average)
Eg1(categorical): 20% of the movies that user U likes(0/1) have Julia, profile for U will have 0.2 in component Julia.
Eg2(numerical): User U has average rating 3, there’re 3 movies he watched with Julia as a star, having rating 3, 4, 5, then corresponding Score is (0+1+2)/3=1.
Collaborative Filtering
Similarity of Items is determined by the similarity of the ratings of those items by the users who have rated both of them.
