In a nutshell, AI (Artificial Intelligence OR ‘Machine’ as our users like to call it) recommendation considers both QUALITY of an article and user engagements.
Below is a simplified version of how the system works.
First, every user has a profile, describing the user’s interests, in terms of keywords, topics etc.
For an article, after it passes review, AI analyzes the content and identifies keywords, topics etc. AI also provides a quality score. Articles with good writing and expertise get a higher quality score. Articles that are likely vulgar/porn, or clickbait get a lower quality score. The quality score is also based on account quality, whether the account has consistently produced good quality articles.
When a user comes in, AI finds all the articles that potentially match the user’s interests, and picks the top ones using the engagement score. The engagement score is based on:
* CTR (Click Through Rate) – How likely will a user click on this article, given the title and the picture.
* Reading time – How long will the user read the article.
* Share – How likely will the user share the article.
* Favourite – How likely will the user mark it as favourite or save for later.
There are additional boosting or demoting logics, combined with the engagement score.
*Articles from subscribed (followed) authors will be boosted (will be shown to many users)
* Articles classified as high quality will be boosted.
* Articles with many comments will be boosted.
* Articles that are newly published will be boosted.
* Articles classified as low quality will be demoted (‘shown to fewer users) such as likely vulgar/porn, or clickbait.
* If there are multiple articles with similar content, only the one with the highest-ranking score will be picked, the rest will be filtered.
* Once an article reaches high impressions, a thorough 2nd review is conducted. If the article is deemed missed by 1st review, the distribution will be stopped.