Thanks for testing! I'll write a description of how it makes the predictions and add it to the tool.
But in short, it uses a machine learning model that I trained with a dataset that contains all stories and comments between 2006 and 2017: https://www.kaggle.com/hacker-news/hacker-news
I've tested various approaches, and currently, the algorithm takes the title as an input and transforms it into an array of numbers between 0 and 1 (each character is a number). Then I give these arrays to the machine learning model (brain.js feed-forward neural network) and the number of scores as an output. After learning and iterating over the data, it spits out the prediction model that I can use to predict the outcome of different title variations.
I've tested the algorithm with approx. 10.000 posts and it has been able to predict 60% of the cases correctly. So, it's not perfect yet, but I use this method in a situation where I don't have any experience of which type of title would work + I don't have time to do "proper" pre-testing.
Thank you for this tool! I think it's a cool idea. Also, it is private, for not sending any requests.
What I would love is the actual predicted number of points (I don't mind that the error would be quite large given the title alone, given that it is not overfit).
With ”Biden sucks” you could expect something more interesting in the context of hackernews. ”Trump sucks” is something so obvious and whole heartedly agreed in hn that it is a blasé/tldr. :)
I tested with: Title A: Google pledging to spend 2 billion by 2022 to get rid of php on the web.
Title B: I stopped using php and here’s how it saved my marriage.
Title B won.
Note: nothing against php, just wanted to test some typical cringey, hyperbolic tech article titles I constantly see LOL.