The science behind online content recommendations

Experts explain how the application of algorithms, large data sets and human curation are making predictions even smarter

Published in Guardian Media

'Thumbs up': a view of Pandora's in-app content recommendation system.

Have you ever wondered how online providers of good and services decide which products, songs, books or movies they suggest to you? It may seem like magic but behind the scenes there's a lot of data being mined, complicated algorithms at work and many researchers who have been striving since the early days of the web to figure out ways to more accurately predict what you like.

Organisations like Amazon, Pandora radio and Netflix can only make guesses about what you might like based on the data available to them, typically from your use of their services and any other data they have shared by partners such as advertising networks. The recommendation system used by Amazon is called affinity based item-to-item collaborative filtering. Within a moment of my loading the Amazon website recently, a message appeared on screen, stating that Terry Pratchett's newest book was about to be released as a paperback, based on my previous purchase history of his books in paper and Kindle ebook format. Personally, the most effective suggestion system I've experienced is theMusic Genome Project (MGP) used by the freemium online music service Pandora. Pandora has over 200 million registered users collectively in the US, Australia and New Zealand, who listened to 1.49bn hours of music in March 2013. Since the company started in January 2000, users have personalised the music they listen to by making over 25bn thumb ratings. These are used in conjunction with each song's MGP metadata to make suggestions that are unique for each individual user.

Jane Huxley, Pandora's Australia and New Zealand managing director, explained to me that Pandora is based on the belief that "each individual has a unique relationship with music – no one else has tastes exactly like yours. So delivering a great radio experience to each and every listener requires an incredibly broad and deep understanding of music".

MGP has been built over 10 years by a trained team, a typical member being a professional musician who has completed several years of tertiary education in music theory, composition or performance. Each song added to the Pandora catalogue is analysed with up to 450 distinct musical characteristics by a member of this team, not a machine or other automated process.

Pandora is great at surfacing surprising "unknown unknowns" – songsthat delight you once you hear them that you would probably never have discovered if the MGP hadn't suggested them to you because they share underlying characteristics with other songs you like.

I also spoke to chief content officer at the international device and eBook vendor Kobo, Michael Tamblyn. In his view "recommendations aren't just nice to have. They are critical to help a reader to find books that will interest them".

In Tamblyn's opinion, suggestions can and should be more insightful and surprising than "people who bought this also bought that". In his vision of an ideal future, the Kobo store would shape itself to the reader, constantly learning about their tastes, adjusting and surfacing new titles based on what they've liked.

Tamblyn said that "great recommendations, whether they be in person or on a device, are always the result of sifting huge amounts of information about which books have sold, when and to whom. A great human bookseller does this intuitively – thinking about previous purchases, trends, what's selling, what friends are reading, what you like and don't like. We do the same thing using data, analytics and algorithms, social activity, ratings and other data to suggest your next great book".

What about the future? Cxense has created a context aware software as a service recommendation system that is integrated into the platforms of their customers. Cxense goes beyond similar product-type recommendations to examine nuanced factors like what time of day a website is being visited, what device is being used to access it and the prior searches a user has made that reveal intentions related to the page they are viewing.

According to Mark Pritchard, senior vice-president of engineering at Cxense, if a news website reader was examining an article about David Cameron, Cxense could recommend other articles related to things mentioned in the article using the context of the reader having an interest in British politics, the EU, a particular industry affected by government policy, etc.

Pritchard thought that sentiment-based analysis such as a mood-based music player would be pretty easy to do if music service users opted in to data mining of the words and phrases they used on social media platforms like Twitter and Facebook. For example, if a user had tweeted "going out for a run" in the last 15 minutes, the suggested music could be faster in tempo.

In the future as recommendation systems further improve their accuracy in offering customers more relevant content, website operators will have to walk a fine line to avoid incurring a backlash of complaints for being creepily accurate.