Introduction to Recommender Systems — How Social Media Already Knows What You Want
What is a recommender system?
If you are an internet user, Google has likely had a tremendous impact on your life. Returning over 6.48 billion results every day to users across the globe, Google is by far the most popular and powerful search engine in the world (99 Firms, 2020). Google has enormous control over what content the vast majority of us internet users see on a day-to-day basis. When we have a query we believe the internet can solve, we nearly always place our trust in Google to figure out what we want, and to transport us to whichever corner of the internet possesses the answer we need, all in the blink of an eye. Considering the search engine performs this task over 78,000 times a second, the power of the Google search engine is not to be understated.
The search engines of Facebook, Instagram and LinkedIn perform a similar task, although these engines tailor their results toward the user far more than Google. For example: if you live in Ireland and search for “John Smith” on Facebook, the search algorithm must recognise that you are far more likely searching for your neighbour John Smith, rather than a John Smith living in New York. The equivalent goes for searching “St Patrick’s Day parade”, or “cat photos” for example. Here, Facebook uses what is called a “Recommender System”, a filtering system which attempts to provide the user with content which is most relevant to them and is not just limited to search engines. Google, Twitter, Netflix and even Tinder use similar systems to try and predict accurately the kind of content you want to view.
You can see recommender systems in use not just in search engines; most social media apps & websites try to grab your attention with personalised content the moment you open them up. When starting up Netflix, or logging into YouTube, you are immediately greeted by boundless videos which the recommender engine determines you would enjoy, based on what you have watched previously. Instagram and Facebook load your feed with posts from users or pages you tend to like, and dating sites such as Tinder recommend matches based on your preferences from the past. How do these sites achieve this? How does a computer algorithm figure out the kind of things you like and predict what you want to see next, even before you may have thought of it yourself?
Ways to recommend
The key lies in the system being able to determine if two pieces of content are similar or not. There are two traditional methods of filtering similar material from dissimilar material: “Collaborative Filtering” and “Content-Based Filtering”. Collaborative filtering compares decisions you have made with similar decisions made by other users to build a model of what you might like. Content-based filtering tries to match characteristic tags attached to content to recommend content with similar qualities (Melville & Sindhwani, 2011).
As you search through the millions of products on Amazon, you’re likely to see something along the lines of “Customers who bought this item also bought…”. This is a prime example of collaborative filtering. Using data from others’ shopping history, Amazon is trying to predict what you may want to try next.
Content-based filtering, on the other hand, you can see more clearly on video sharing sites, where videos are tagged by the type of content they show. “Where do black holes come from?” would be tagged as educational content, while “TOP 10 EPIC TRICK SHOTS” would probably be tagged as sports content. Or maybe gaming, it’s hard to know these days. Sites such as Netflix would use a movie or series’ genre as a characteristic tag. This method is regularly used where it is more complicated for a system to determine the nature of the content on display, e.g. for videos.
What about the future?
Companies are now trying to move to a more hybrid approach of system, combining both collaborative and content-based filtering, as well as some other techniques, to compile a list of recommendations for you. Take YouTube, for example: creators on the site tag their videos as entertainment, sports, gaming, or whichever topic suits their content best (content-based filtering). Part of the YouTube algorithm also takes into account which videos you like or dislike, recommending content which users who have liked similar content to you also like (collaborative filtering). This is why nearly all YouTubers invariably insist you “smash the like button, ring that bell and subscribe for the latest content” — at least, that’s the mantra that has been stuck in my head for the last number of years. In doing so, you increase that YouTuber’s probability of being shown on your home screen by their high-tech, hybrid recommender system (Melville & Sindhwani, 2011) (Adomavicius & Tuzhilin, 2005).
Hybrid recommender systems are the future for businesses who want to predict their users’ likes and dislikes — knowing what kind of content you like and knowing what people similar to you like is the perfect recipe for predicting what you may want in the future. The power of companies to figure out our interests from our experiences some may say is intimidating and intrusive, a debate for another day which I do not believe will end anytime soon. I hope that now at least, you have a better understanding of how companies use your information to predict your internet preferences!
References
· 99 Firms, 2020. Google Search Statistics. [Online]
Available at: https://99firms.com/blog/google-search-statistics/#gref
· Adomavicius, G. & Tuzhilin, A., 2005. Toward the Next Generation of Recommender Systems: A Survey of the State-of-the-Art and Possible Extensions. IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, 17(6), pp. 734–749.
· Melville, P. & Sindhwani, V., 2011. Recommender Systems. In: C. Sammut & G. I. Webb, eds. Encyclopedia of Machine Learning. New York: Springer Reference, pp. 834–842.