Collaborative filtering

All unapproved Citizendium articles may contain errors of fact, bias, grammar etc. A version of an article is unapproved unless it is marked as citable with a dedicated green template at the top of the page, as in this version of the 'Biology' article. Citable articles are intended to be of reasonably high quality. The participants in the Citizendium project make no representations about the reliability of Citizendium articles or, generally, their suitability for any purpose. This article is currently being developed as part of an Eduzendium student project. The course homepage can be found at CZ:Special_Topics_2010. To provide students with experience in collaboration, you are warmly invited to join in here, or to leave comments on the discussion page. The anticipated date of course completion is 13 August 2010. One month after that date at the latest, this notice shall be removed. Besides, many other Citizendium articles welcome your collaboration!

Definition

Collaborative Filtering(CF) refers to the use of software algorithms for narrowing down a large set of choices by using collaboration among multiple agents, viewpoints, and data sources.

Overview

The term Collaborative Filtering was first coined by the makers of one of the earliest recommendation systems, Tapestry. The basic assumption in CF is that user A and user B's personal tastes are correlated if both users rate n items similarly.

Collaborative Filtering systems follow this approach to produce recommendations:
1. Gather ratings from users and maintain user's ratings in a database.
2. Compute the correlations between pairs of users to determine a user’s neighbors in taste space
3. Compute the ratings of these neighbors to make recommendations.

Collaborative Filtering requires ratings for an item in order to make a prediction for it. (Novelty and serendipity,accuracy and coverage).Unlike Content filtering, Collaborative filtering does not require content; instead, CF requires ratings for an item to make a prediction for it.(RS)

Collaborative Filtering Techniques

Collaborative Filtering techniques can be separated into 3 classes:

1. Memory-based(Heuristic) Recommendation Technique

Memory-based algorithms make predictions by operating on data (users, items and ratings) stored in memory. Nearest neighbor algorithms are the most commonly used CF algorithms. Users similar to the current user with respect to preferences are called as neighbors. Nearest neighbor algorithms can be further classified into User-based nearest neighbor and Item-based nearest neighbor algorithms.

User-based nearest neighbor algorithms generate predictions for a given user based on ratings provided by users in the neighborhood. A common approach to user-based nearest neighbor algorithm is:
1. Weight all users with respect to similarity with the current user. Similarity weighting is done by either using:

• Pearson correlation coefficient between ratings for current user c, and another user, u.
  
• Cosine-based correlation where in the two users c and u are considered to be two vectors in an m-dimensional space and the similarity between the two is measured by computing the cosine of the angle between them.
  

2. Select a subset of the users (neighbors) to use as predictors.

• Neighbor selection is done by finding similar users or users having a similarity weighting above a certain threshold.
  

3. Normalize ratings and compute a prediction from a weighted combination of the selected neighbors’ ratings.
4. Present items with highest predicted ratings as recommendations.

Item-based algorithms generate predictions based on similarities between items.
1. The correlation of each item i_k with other items is computed.
2. For each user u_i , the ratings of the items highly correlated with i_k are aggregated.

Advantages of Memory-based algorithms:
1. It is easy to implement.
2. It can detect complex patterns easily.
3. It scales well with correlated items.
4. It does not require the content of items, only the ratings are sufficient.
5. New data can be added easily.

Disadvantages of Memory-based algorithms:
1. It depends on human ratings.
2. Correlations are skewed when data is sparse.
3. Time and memory requirements increase linearly with the number of users and ratings.
4. It cannot recommend for new users and items.