In the 1990s, as hundreds of television channels streamed live programming twenty-four hours a day, seven days a week, millions of people pondered the simple problem of what to watch on TV. Most programming was dreadful and boring; for most users, only a tiny percentage of the content was likely to be interesting. A student at MIT, Nicolas Negroponte, proposed a solution. Negroponte imagined the future of television to be embedded with intelligence agents in every device. “Imagine a future,” Negroponte wrote, “in which your interface agent can read every newswire and newspaper and catch every TV and radio broadcast on the planet, and then construct a personalized summary. This kind of newspaper is printed in an edition of one… Call it the Daily Me”
Today, Negroponte’s vision has become a reality. We consume huge amounts of personalized media and browse personalized retail with the help of top recommendation algorithms. Personalization has become apparent in nearly every aspect of our online lives, whether in your TikTok For You page, YouTube’s recommended videos, or the top products on your Amazon search.
How does YouTube even know what video to recommend to me? And why are their recommendations so good? Currently, three main types of filtering are used in recommendation systems.
Collaborative filtering
Collaborative filtering is one of the most common recommendation systems. In 1990, a team of researchers at the Xerox Palo Alto Research Center (PARC) started tinkering with collaborative filtering, which they ran in a program called Tapestry. Tapestry tracked how people reacted to the emails they received – which items they opened, deleted, replied to, or read for a long time. The PARC team used this information to help other users order their inboxes automatically. Emails that were usually high in engagement would move to the top of the inbox, while emails that were frequently deleted or left unopened would shift down.
Tapestry perfectly displays how collaborative filtering predicts a user's interests by identifying preference information from many other users. This system uses the similarity of user behavior, given previous interactions between users and/or items, to predict a future interaction. The underlying intuition behind collaborative filtering is if persons A and B made similar decisions/purchases in the past, they will likely make similar decisions and purchases in the future.
There are two main types of filtering: item-based and user-based.
Item-based collaborative filtering focuses on the similarity between products by taking data from similar user activity, i.e. how many users that bought item X also bought item Y. If the correlation is high enough, a similarity can be presumed to exist between the two items and recommendations will be made for both product X and Y. For example, suppose most people who purchase bracelets go on to buy necklaces. In that case, bracelets will be recommended for people who buy necklaces, and necklaces will be recommended for people who buy bracelets.
User-based collaborative filtering operates on a system by finding the “nearest neighbor”. The “neighborhood” uses rating patterns of similar users (i.e. the “nearest neighbors”) to give recommendations. The largest caveat of user-based collaborative filtering is that recommendation quality depends on arbitrary user interests. Some users could get better recommendations than others, as they have more members in their neighborhood for the algorithm to work with.
Content filtering
Content filtering uses the attributes of an item to recommend other items similar to the user’s preferences. The math is complicated, but the idea is that similarity is calculated from the item’s features and the user’s preferred features from previous ratings. Content filtering uses the same item- and user-based systems as collaborative filtering but requires developers' more domain knowledge to attribute accurate product features. In contrast to collaborative filtering, other users’ data are not required to start recommendations.
Amazon’s book recommendation feature in the late 1990s is a great example of content filtering. If you’ve spent a lot of time browsing the latest Stephen King novels but only glanced at the new diet guide, you might see more supernatural horror books and fewer health books.
Contextual filtering
Contextual filtering uses users’ contextual information in the recommendation process. Systems acquire data from information like location, time, device, date, etc. The sequence of contextual user actions will be applied to the current usage context to predict the next action. Many hybrid models are made to apply contextual awareness to collaborative filtering for optimal results.
In Google searches, contextual filtering plays a role of paramount importance. For example, when a device from the location of an African Safari tour searches for “panthers”, they are probably looking for wild cats, whereas a device from North Carolina at 8:00 pm on a Sunday night searching for the same phrase probably means the Carolina football team.
Works Cited
Bostrom, P. and Filipsson, M. (2017). Comparison of user based and item based collaborative filtering recommendation services.
Educative: Interactive Courses for Software Developers. (n.d.). What is content-based Filtering? [online] Available at: https://www.educative.io/answers/what-is-content-based-filtering.
Minds, C. (2020). What are the top recommendation engine algorithms used nowadays? [online] Medium. Available at: https://itnext.io/what-are-the-top-recommendation-engine-algorithms-used-nowadays-646f588ce639.
NVIDIA Data Science Glossary. (n.d.). What is a Recommendation System? [online] Available at: https://www.nvidia.com/en-us/glossary/data-science/recommendation-system/#:~:text=A%20recommendation%20system%20is%20an.
Pariser, E. (2014). The filter bubble : how the new personalized web is changing what we read and how we think. New York: Penguin Books.
Comentários