14 September 2016 397 words, 2 min. read

Recommendation algorithms : how big is the Filter Bubble in actuality

By Pierre-Nicolas Schwab PhD in marketing, director of IntoTheMinds
Recommendation algorithms have been accused to trap users in a filter bubble, to promote the balkanization of information and hence to reduce serependity. The biggest opponent of recommendation engine is probably Eli Pariser who invented the very term of “filter […]

Recommendation algorithms have been accused to trap users in a filter bubble, to promote the balkanization of information and hence to reduce serependity. The biggest opponent of recommendation engine is probably Eli Pariser who invented the very term of “filter bubble”.

Despite very convincing arguments in his 2011 book, Pariser somewhat failed at backing his theory with figures (which is perfectly understandable given the novelty of this topic back in 2011).

What we propose to do in today’s post is to present you the results of one scientific study that looked precisely into this. The results may not be as you may expect.

The setting of the study

The study is based on MovieLens’ recommendation engines and in particular on the activity of 217,267 unique users since 1997. Some 20,000 movies were rated and 20 million ratings given.

What was studied

Two groups of users were studied : those who followed recommendations provided to them by the recommendation algorithm (“followers”) and those who didn’t (“non-followers”). The authors explored the diversity of the content consumed using the “tag genome data”. We won’t get into more details about this as it may be relevant only for only a small part of the readership. If you are a data scientist wishing some more info, drop us a line and we’ll be happy to answer.

Let’s now move on to the results.

Results : does the filter bubble really exist ?

The first conclusion is that content diversity decreases slightly over time even for users not following recommendations. It seems therefore that not exposing users to recommendations doesn’t prevent them to narrowing down the content they consume.

The second conclusion is that the diversity of the movies watched decreased more for “non-followers” than for “followers”. That’s a very counterintuitive result

The third conclusion is that the followers were more satisfied by the movies they watched than the non-followers. Over time indeed, the ratings given by followers were better than ratings given by non-followers, suggesting that the recommendation engine did its work and was able to propose the right stuff.

Conclusion

This study contradicts the existence of the filter bubble and shows actually that there is a slightly positive effect of the recommendation algorithm in discovering more diverse content.

However the effect is only slight and much research is still needed to turn recommendation engines into exploration engines promoting curiosity and serependity.

Image : Shutterstock



Posted in big data, Innovation, Marketing.

Post your opinion

Your email address will not be published. Required fields are marked *