Large distances silently ignored

[thomasp85]

When a point is so far away that the exponential to the negative distances returns 0 for all distances, the d2a function silently fails and leave the affinities for that point at zero. This is due to a lack of NaN check of H.

I'm working on an R implementation and what I do is set beta[i] to beta[i]/10 in the case of H == NaN and continue calculations.

[AdrianP-]

Yes, I dealt with that. But in my opinion is better option normalize the input.

[jeroenjanssens]

Hi @thomasp85 , it's been a while :) It's great to hear that you started working on an R implementation. Did you ever get a chance to finish it? How's that trick you propose working out?

[thomasp85]

2 years fly by :-)

Yeah I made a working R implementation but never got around to releasing it (it got caught up in dreams of a coherent system for running outlier detection). I should just release it...

As far as I remember the trick did the trick:-)

[thomasp85]

Now it is on GitHub at least :-)

https://github.com/thomasp85/anomaly/

[jeroenjanssens]

Thanks @thomasp85 for the suggestion.

By the way, I really like that you ported SOS to R. If you plan to continue with it, then a mention to the original paper would be much appreciated.

[thomasp85]

Sure - I'll cite all relevant sources. Currently it's not near any publishable state but once I get to finish it you'll get credited properly