A Quick Sketch of a Human-driven Paper Recommendation Mechanism

This note is about the overall system by which we hear about papers and give them attention. The focus of the note is a very rough draft of a constrained human-based recommendation mechanism.

Update (Dec 5, 2023): this mechanism is now being implemented at recnet.io.

How we hear of papers to read has changed dramatically in recent years. During my PhD, I would scan the proceedings of conferences. Later, I followed arXiv feeds, and even later looked for recommendations and paper announcements on Twitter. All these ways worked for some time, but became outdated, intractable, or otherwise just devolved into useless dynamics. Right now, I am a bit at a loss.

There are numerous attempts to fix this problem with recommendation systems.¹ I never found these particularly useful or compelling. I am sure many others find them much more useful. But, being surrounded by recommendation systems I find of little utility, I just don’t see that the solution is in automation. Maybe I’m just old.

I occasionally tried to think recently how would a good human recommendation mechanism look like. So, here’s a concrete suggestion. It starts very similar to contemporary social networks, but it’s designed to be improvished in certain ways and contain information bottlenecks that increase communication cost. This is intended to limit the amount of time the system consumes from its users, while increasing the quality of information passed. The mechanism aims to piggyback on (a) the reputational incentives of academic research, and (b) the utility of shaping the mindset of your research peers (i.e., influence has a lot of value).

The mechanism is specified as follows:

• Users have to sign-up and follow other users to receive recommendations from them. Users can see how many people follow them, but they can’t see who is following them.

• The process works in weekly cycles (i.e., seven-day periods), with a fixed cutoff time (e.g., Tuesday 00:00 GMT).

• Each user may recommend one paper per cycle. They may change the paper they recommend each cycle as many times as they want during the cycle (i.e, before the cutoff time).

• A recommendation is made out of (a) a link to the paper, (b) paper title and authors, (c) a very short tl;dr message of 280 characters.

• Recommendations are released once a week, and are sent via email immediately after the cutoff time. Each user receives an email with a list of recommendations from all the people they follow. The list is shuffled, independently for each email sent.

• There are no options to comment on recommendations, or to like/dislike them. The communication channel is one-sided.

• Following the cutoff time, the process starts anew.

• Public user profiles list all past user recommendations. Past recommendations can be deleted by the user. User profiles do not list how many people the user follows, or who they follow.

• [still considering:] User profiles list the number of their followers.

• [still considering:] Users may not recommend their own papers (enforcement is tricky), or may recommend one paper of their own during each period, but that will be marked separately and won’t use their weekly recommendation budget.

1. Examples include Semantic Scholar’s Research Feeds, Google Scholar’s recommended articles, and many more.