Relation to hazardous by probabl - is that a copycat project?

[fkiraly]

@CamDavidsonPilon, may I ask what is going on with the hazardous library by the company probabl (French company close to some sklearn developers)?

The most advanced state is probably the "ICML 2024" branch https://github.com/soda-inria/hazardous/tree/icml-2024 On the technical level, it seems like:

• it wraps lifelines without clear credit, see example 1, example 2, and some other packages like pycox
• it duplicates some features from lifelines, relying also on heavy wrapping, see metrics module
• it it incomplete, has not been released on pypi, has not been worked on in the public repository since ca half a year. It may be in stealth mode, although if current progress can be extrapolated, it wlil not be ready in a state comparable to lifelines, for one or two years.

On the publicity level:

• it is prominent on the probabl webpage, listed as one of the 8 major libraries that probabl is funding (see below)
• it is currently being advertised on linkedin
• the name of the branch "ICML 2024" seems to imply that there will be a major fesature at ICML 2024, July.
• Olivier Grisel and probabl are repeatedly placing survival modelling content on social networks

I find that odd, especially the statement that it is one of probabl's flagship projects, given the basically completely unready state of hazardous and how much it wraps lifelines.

Do you have an agreement with probabl?

probabl webpage: https://probabl.ai/open-source

Screenshot, retrieved June 9, 2024

[CamDavidsonPilon]

Hi @fkiraly,

Thanks for bringing this to my attention. We don't have any connection to that project or company. While it would be nice to receiving trickle-down funding from a VC-backed company, I'm not expecting it. Nor am I expect expecting any attribution beyond import statements.

What I do expect is that lifelines is a placeholder. These devs probably want to rewrite important models themselves (they have more experience with Cython and probably want control over internals) rather than defer out to lifelines. Probably they used lifelines-wrapped models for early demos.