Recommender Systems - Introduction
Importance of Recommender Systems
Over the past few decades, with the advent of Amazon, Netflix, Youtube and many other similar web services, recommendation systems have taken up more and more space in our lives. From e-Commerce (offering customers articles that might interest them) to online advertising (offering users the right content that matches their preferences), recommendation systems are now unavoidable in our daily online travel. Huge conferences are hosted in different parts of the world annually, featuring many big companies. The RecSys conference is a good example of this. Recommendation systems are really important in some industries because they can generate huge revenue when they are effective, as well as being a way to stand out significantly from competitors. Here follows some of the most common general techniques used in recommender systems.
Collaborative Filtering
The main idea of collaborative filtering is to search for information or patterns using collaborative methods between multiple data sources. Data sources usually consist of users and elements such as movies, songs, etc. These data sources mainly consist of users and items that users can interact with (purchase, click, watch). The search for these patterns is performed by collecting and analyzing a huge amount of data about the behavior and activities of users, as well as items. Collaborative filtering is one of the most common general methods for recommender systems being applicable within several big use areas. Collaborative Filtering is frequently used in social networks such as Facebook, LinkedIn, mySpace, Twitter, etc to effectively recommend new friends, pages, groups as well as who to follow or what to like. But also applications such as Youtube, Reddit, Netflix, etc make use of collaborative filtering. Collaborative filtering is viable in all applications where one can observe connections between a user and their registered friends or followers. But it is also widely used within e-commerce, where Amazon is a good example who popularised algorithms for item-to-item based collaborative filtering.
Content-Based filtering
Another well known method when implementing recommender systems is content-based filtering. Content-based filtering is usually used for movie recommendations or in other environments where, for example, keywords are used to describe the items of the system. Among the actors of recommender systems using content-based filtering one can find The Internet Movie Database and other popular services for movies. The methods of Content-based filtering usually use the correlation between item features and preferences from a user's profile, in contrast to the collaborative filtering approach that selects items based on the correlation between users with similar profiles. For the above to work, the system needs to deploy some learning technique such as Bayesian networks, clustering, decision trees, neural networks, reinforcement learning, Nearest Neighbour etc. The system uses these techniques when observing historical data of users to learn their preferences. The intention is that, after sufficient amounts of data have been observed, the system should be able to predict future behaviour of a specific user
Hybrid Systems
A hybrid approach to implementing recommender system means designing the system so as to making use of several recommendation techniques such as for instance collaborative filtering techniques and content-based filtering, making predictions from the combined conclusions of the two. This can be done by unifying the two techniques, by adding features from one into the other or simply by running algorithms for both techniques separately and then combine the results in some way. The most common example of a hybrid-based recommender system is the one used by Netflix. While an environment like Netflix is well suited for a hybrid recommender system, it does not fit everywhere. Why Netflix is considered a hybrid system:

  • Collaborative Filtering: Observing the watching and browsing habits of similar users.
  • Content-Based Filtering: Observing users with equal preferences and how they rated certain movies.

© All Right Reserved. Columb Labs Inc.
e-mail us: info@columblabs.com