Date of publication: 2017-08-23 10:28
If we want to have the both content and collaborative filtering's good worlds, then we could adopt a hybrid approach. They could be combined in the following ways:
Traditionally, this questions are answered with peer recommendations(word of mouth, forums, blog posts or reviews) or expert advice(columnist, librarian and recommendation of someone who has domain expertise). Traditional methods are good but limited in the observation spaces that recommenders have to begin with. Your peers could only read so many books, could only visit so many restaurans, could only watch so many movies. Second, they are biased towards their preference(naturally) where when you want to make a decision, you want to be biased towards yourself in order to maximize the decision outcome. Third, a person may not have access to these traditional methods. She may not have peers who share somehow same taste in music and movies, she may not have access reading expert advices.
User 7 watched Copycat , Powder , Now and Then , It takes Two , (( Crime | Drama | Mystery), (Drama | Fantasy | Mystery), (Comedy | Drama), ( Comedy | Family | Romance ). and our collaborative filter recommends Volver (Comedy | Crime | Drama ). Not so bad, huh?
This subsection of machine learning methods also have connections with information retrieval and actually the problem could be formulated as an information retrieval problem as well. Consider Google, the links(items) are brought to the first page to the users based on query information, location, user history and so on. Therefore, most of the algorithms invented in information retrieval could be adopted to recommendation systems with minimal changes. The reverse may not hold true in general, though.
Difference business and product needs and a variety of algorithms that could be used for recommender systems yielded a rich set of methods that could be used for recommendations.
Due to these shortcomings, computer based recommender systems provide a much better alternative to the user. Not only they do not have these shortcomings of the traditional methods, but also they could mine the historical information of the user and demographics information which may result in a more accurate and finely-tuned recommendation for a problem that what traditional methods could offer.
Note also that, in a production system or a website the prediction rating(or confidence) may not be meaningful at all. Therefore, although it is important to have confidence for the recommendations that will be displayed to the user, the
Recommender systems is a family of methods that enable filtering through large observation and information space in order to provide recommendations in the information space that user does not have any observation, where the information space is all of the available items that user could choose or select and observation space is what user experienced or observed so far.