A smarter Headlines Today
The goal of Headlines Today is to keep users updated and informed as quickly as possible and to have an app in which users know that every story they read is reliable, has been checked and is not fake.
With this latest version of Headlines Today we have taken this to the next level. All stories now appear with full images, as, to paraphrase Henrik Ibsen, a picture is worth one thousand words.
You can now subscribe to channels and topics. We have updated our system so you can subscribe to anything. You could subscribe to the word “the” and see every story with the word “the” in it (not that you would want to). You may have noticed you can both see and subscribe to COVID-19 news. This is an example of this feature.
We are currently working on a meaningful topic extraction system. Our system can already extract named entities, but this is not useful for topics such as “Brexit” or “Coronavirus”. For this we are using Latent Dirichlet allocation.
On previous versions of Headlines Today, articles would be downloaded when they were opened. This resulted in a slow and therefore poor user experience as every article was a friction point. We have upgraded our (what we call) our HTBack end to a 2 cluster Raspberry Pi (simple, yet brilliant at the same time). These can download 2000 articles in about 15 minutes where one is downloading the articles, sending the result to the other which is parsing and uploading them. This means that when a user opens an article, it is already downloaded and ready to display. On the client, we make use of on device machine learning models which predict how many articles a user will open and download them. That way we are minimising the amount of data a user uses.
Headlines Today can now recommend sources for you. This is a big step forward. On previous versions, the user would be asked what sources they liked on launch. This is an easy way to do things, but if a user does not want to spend time choosing from a wide range of sources, it is not optimal. It is far easier to recommend sources and then give them the option to personalise. Of course, since a user’s subscriptions are personal to them, this is all done on device. Over time, we will start localising the model to ensure that people get more relevant recommendations. Read our post on ml with V12 to see more.
We will continue to post about our updates on Headlines Today. Check out the posts mentioned above for more information on this new release.