Machine learning will save you from the flawed Follow friends

Instead of capitalizing on your friends poor taste in music and news, lets learn about you and how you behave, read, listen, all to bring you the best experience possible. Both Spotify and Pocket are good examples of companies who have done this right.

On oh so many services you do find the “Find My Friends” button that will help you get better content in service you use. You find those Facebook and Twitter logins everywhere that want access to your friends to bring you a better “experience”.

I find it mainly a flawed system. This due to the simple fact that the people I follow on Twitter and the friends I have on Facebook etc, might not be the ones I share interest with on the new service. I’ll just give two examples:

  1. Spotify. Spotify launched it’s social discovery feature a long time ago and did it together with Facebook. The thing is that my Facebook friends do not listen to the music I want to discover, they are mainly my friends on Facebook for other reasons.
  2. Pocket. Pocket is my favorite “read-later”-app. It just launched a follow friends feature, I tried it out the only result was that I found the same links that I had already seen on Twitter, from the same people that I follow on Twitter and now in Pocket as well.

There are many many more examples. There are occasions when this do work though. I love both the Newsle and Nuzzle services as examples. Newsle finds article and posts about my connections on LinkedIn, but also from my mail-contacts. Nuzzle simply digests the most intereseting posts from my network from a specific time based on shares etc within my friend / follower network, I do have Nuzzle deliver a daily digest every morning with the last 24 hours hot news in it.

Is there a solution to this flawed “Find my friends” thing?

Naturally there is. I actually took both the above examples since they have alternative ways or actually have changed there service since it launched.

The solution is to know the individual better , learn from the user and then from that, use the data at hand to discover top content for the user.

Spotify

This was for long a disaster and discovery was the Achilles Heel of Spotify (in my opinion). Spotify realized this and did two things. Spotify acquired Echo Nest and they set-up an additional internal machine learning team. The result speaks for itself. Today the Discover Weekly is a huge success and after 10 weeks since launch Spotifys Discover Weekly had streamed 1 billion songs.

Spotify naturally uses its own technology as well as the acquired Echo Nest platform. it uses natural language to understand blogs, titles and meta data. Then there are many other machine learning things as well as Kafka to work with the data in real-time.

Compared to the the list of what my friends on Facebook listened to, this is huge progress and today a big portion of my listening comes from either curated or automated playlists on Spotify. My user behaiviour have completely changed since Spotify started to work dedicated with Discovery.

To read up on the Discovery features in Spotify:

Pocket

So I am not a fan of Pockets “What friends share” in the Recommended section, but I am a fan of the posts that Pocket itself recommend to me. It is based on my reading and then presents personal recommendations for me. It is a much simpler use case then the above Spotify one. Pocket also uses natural language processing and simply uses the IBM Watson AlchemyLanguage API on its content. They look at my content and let the Watson service digest entities as well as extract conceps from that content and then use those to find other popular content that matches that info. The wast majority of recommendations I get from Pocket in the Recommendation section (not the ones shared by friends I follow) are spot on and I would say that about 50% of my saves in Pocket has originated from that part of the feature.

For me it is not about what others want to read, it is what I want to read. It took the Pocket team only a very short time to implement as well. There is a case study available to read some more (PDF).

 

 

Leave a Reply