It's less harmful than people think. In a "collaborative filtering" or "chronological timeline with boosts, retweets, etc." world news gets filtered for riling people up and being inflammatory. You might get the conclusion that news itself is a bad thing.
If you subscribed to the RSS feeds of 80 publications and picked articles at random you might find 2-3% of the articles are inflammatory. Filter by personal interests and you might get all kinds of weird stuff but it won't be hateful.
I’ve got YOShInOn which is similar to that, it ingests about 2000 articles a day and picks 300 to show me and uses a BERT-based classifier and clustering system.
It looks like TikTok for text or the old Stumbleupon (Somehow completely forgotten in a very short time) but it uses content-based filtering instead of collaborative filtering. I’ve thought about applying bandits, sequential recommendation and such but stuck with what I know because I know my scheme is producing a more-or-less valid sample for evaluation.
I worked on a text classification project in 2005 that made me wonder why RSS readers kept failing with the same failing interfaces that were always failing. It took the insanity at twitter for me to pull the trigger on it this december.
Super interesting, will follow your work. Other media could probably be integrated in a similar capacity or from the embeds in the article. Thanks for working to save the signal of human thought
4 comments
[ 3.1 ms ] story [ 22.2 ms ] threadhttps://www.amazon.com/More-You-Watch-Less-Know/dp/188836380...
If you subscribed to the RSS feeds of 80 publications and picked articles at random you might find 2-3% of the articles are inflammatory. Filter by personal interests and you might get all kinds of weird stuff but it won't be hateful.
It looks like TikTok for text or the old Stumbleupon (Somehow completely forgotten in a very short time) but it uses content-based filtering instead of collaborative filtering. I’ve thought about applying bandits, sequential recommendation and such but stuck with what I know because I know my scheme is producing a more-or-less valid sample for evaluation.
I worked on a text classification project in 2005 that made me wonder why RSS readers kept failing with the same failing interfaces that were always failing. It took the insanity at twitter for me to pull the trigger on it this december.