search

blakestad / recommender

ArXiv mathematics paper recommender
https://www.cliffblakestad.com/projects/recommend
0 stars 0 forks source link

A version which only uses title #1

Open Rubiel1 opened 3 months ago

Rubiel1 commented 3 months ago

Hi, I am interested in your algorithm as a classifier that assigns a subject to a title. By comparing with the math genealogy project, what is the performance of your classifier when the user only provides titles?

blakestad commented 3 months ago

Hello,

Firstly, I think you are talking about my subject identifier, not this recommender.

I have never run a large test based on titles, let alone against the titles in the math genealogy project. It seems like a nontrivial amount of work just to collect that data for testing against in the first place. Just looking around now, it doesn't even look like most listings in the genealogy project have expressed subject classifications.

In general, when tuning the parameters I did set for the classifier, I mostly did so using examples of both titles and abstracts, so I don't have a very good sense of how well it works with titles alone. Nothing has been rigorously tested in terms of the quality of the assigned subjects. I have just tried a bunch of papers I was already familiar with and theses of other grad students I knew and the anecdotal evidence seemed pretty good overall.

Note also that my code can generate more than one classification for a given input (part of the point is that many papers have multiple classifications), which may or may not be what you are looking for.