ArtsAutosBooksBusinessEducationEntertainmentFamilyFashionFoodGamesGenderHealthHolidaysHomeHubPagesPersonal FinancePetsPoliticsReligionSportsTechnologyTravel

What is Collaborative Filtering?

Updated on June 7, 2012

Introduction

Recommendation Systems are a widely accepted tool that can help users search, sort, filter and classify information. They incorporate recommender algorithms which generate personalised recommendations for users. The algorithms use different techniques to build user profiles and find patterns in the data.

The personalisation process has been incorporated into many well known e-commerce sites, and there are a number of those sites which use collaborative filtering methods to generate personal recommendations for users, including Amazon, GoogleNews and Facebook.

Collaborative Filtering

Collaborative (also known as Social) Filtering is an information filtering technique which uses individual user profiles and finds connections between them. The user profiles are constructed using subjective user or item criteria, such as ratings, the time spent viewing a page, and purchase history.

Recommendations are generated for a user using the data in other users' profiles as a pooled resource. The process is modelled on how people innately interact with each other i.e. sharing opinions. For example, if your close circle of friends have all seen the latest movie release and expressed a dislike for it you may not spend the money on going to see it. However, it may be the case that none of your friends like the same type of movies as you do. Fortunately advances in artificial intelligence and the Internet mean there is no restriction to our circle of friends when we want to find something of interest.

It is possible to visualise a simple Collaborative Filtering system as a matrix. The table below consists of ratings for 6 movies by 4 users; the rating scale for this example is 1-to-5 with 1 representing dislike and 5 being a very good rating. The '-' symbol represents a no-vote; either the user has not rated this movie yet or has chosen not to rate it. Any item not rated by the current user is a possible candidate for recommendation.

 
The Godfather
Pulp Fiction
Forest Gump
Toy Story
WALL E
Babe
Harry
5
-
-
5
-
5
Larry
2
-
5
-
4
-
Mo
4
5
2
5
5
5
Donna
-
5
-
5
-
-
Example Collaborative Filtering User / Movie Rating Matrix

User based Collaborative Filtering

If recommendations were required for Harry, the system could employ a user-based approach and use the information in the rows of the matrix. Mo would be selected as the most similar user, because Harry and Mo have co-rated 3 movies and expressed similar preferences about them. An appropriate recommendation for Harry would be "Pulp Fiction" as Mo has rated this highly; likewise a poor recommendation would be "Forrest Gump" as Mo did not like this movie.

Item Based Collaborative Filtering

Collaborative Filtering can also be used to find similarities between the information in the columns of the matrix and make predicitions about what users would like, depending on how similar the movies are, based on the items' rating patterns. For example, there is a perfect correlation between the ratings for the movies "Pulp Fiction" and "Toy Story". Therefore, a good recomendation for Harry might be "Pulp Fiction" as he likes "Toy Story". Similarly, "Babe" would make a good suggestion for Donna.

Are Recommender Systems Worth It?

Do you find the recommendations generated by e-commerce sites worth divulging your private information for?

See results

Comments

    0 of 8192 characters used
    Post Comment

    No comments yet.

    Click to Rate This Article