The targeted users are the paying NYT subscribers or even the people who have their email linked (so that the NYT can drum up patronage).
The goal of the NYT’s recommender system is to curate personalized content for its users in a way the 1. retains subscribership and 2. maximizes engagement with its subcribership.
Maybe this is something that NYT already considers, but I did not hear anything about regionalized content. Maybe that regional-specific content consideration is less important, but I think it might be important to send subscribers regionally-relevant information. I say this because frequently websites will ask if they can have access to your location (and as someone who does not read the NYT, I suspect the NYT asks for it as well). Maybe this would be less reliable as most people, including myself, opt for the no-location- tracking option. However, this may be a useful bit of data to leverage, if it is not covered through connections made to “similar users.” (Though I do not know how VPNs may impact location accuracy)
My first thought as I was reading this article is that the topic analysis seems a lot like the work we did for our third project and the work that was supposed to be done for the sentiment analysis project (admittedly, I did not do that work). So, I believe what happens is that likely folks at the NYT identify articles that qualitatively focus on a certain topic (or multiple topics) and then feed those articles through some sort of text-analysis process that identifies key/important words that occur across the topic (signifying that a certain word may be associated with a topic). Not all words share equal significance within a topic and that would most likely be accounted for with some sort of rating system. All this happens separately from users interacting with the articles.
Then the NYT waits for a user to interact with articles. Once a user begins clicking on articles, NYT can start recommending. Through the beginning of a subscriber’s patronage, this model would rely more on a users pre-existing interest than the recommender system (the same-sex vs heterosexual wedding example). However, once the NYT’s system has a pretty good idea what their users intial interests are, then they can begin being recommended articles based on their own preferences and the preferences of users similar to them. So, if you express a decent interest in the Environment, the NYT recommender system will look at other users’ interests who also show interest in the environment, which may lead you to getting recommendations along the lines of politics or science (just as an example).
According to “Article 1” NYT has no way of officially knowing whether or not a user actually read the article (they can see that a user clicked on an article and they can see the scroll depth, but they cannot assess with 100% accuracy if the article was read). So this is another component that must work it’s way into the model.
Article 1 = https://knightlab.northwestern.edu/2016/03/28/a-quick-look-at-recommendation-engines-and-how-the-new-york-times-makes-recommendations/#:~:text=They%20use%20the%20simple%20method,decide%20what%20your%20interests%20are.
I think that utilizing volunteered location may be helpful, but beyond that I cannot think of any further recommendations.