Mode calculation (samples without modes + and multi-mode) about elixir-statistics

[msharp]

Thanks @PragTob

I have based a lot of the behaviours in this package on the Python numpy and scipy libraries.

The scipy.stats.mode() function returns the lowest value when there are multiple values with the same frequency - https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.mode.html

Can you describe how you are using this function so I can understand your requirements?

[PragTob]

Hi there,

truth be told I'm not really using - I was looking at the library when evaluating if it was worth releasing my statistics library (Statistex) and came across this so I thought I might as well report it as I believe it to be a bug.

That said, I'm not the one to argue with numpy and scipy. It's just what I read up on when implementing mode (or fully truthful, what someone else read up on and I questioned when reviewing it :) )

I think it's important to know about a data set that there are multiple modes and also seeing them as they might be very interesting. For instance I use this mostly with benchmarking. Seeing that there are 2 modes but they aren't directly "next" to each other is super interesting.

In the same vein, when no value occurs more than once and then just the smallest value is reported that seems highly unhelpful to me. In benchmarking terms it would have me believe that the fastest run time is the one that occurs most frequently which imo heavily skews the results.

Anyhow, that's just my perspective/what I remember from reading it up back then. If numpy and scipy do it like this I'm sure it's fine and feel free to close this :)

Tobi

[msharp]

Thanks for the background @PragTob

To be honest, I probably didn't think too much about the mode implementation. It's not a statistic I use very often.

Benchmarking is probably better evaluated with percentiles anyway, especially if you have real (floating point) numbers.

There might be scope for another mode function with a different arity which can support multimodal datasets.

[PragTob]

I find all of them useful, but yes benchee also supports percentiles :) (Currently by default shows 99th%, which might be too hardcore maybe 95th% would be a better default, not sure) But yeah, love me some box plots :)