This might be the way forward but I don't think we're going to do this in 0.3.

Bump. Now is the time to do this.

+1

Great idea.

I support this too, but one concern is that the implementation in StatsBase looks like it will probably be painfully slow. Has anyone benchmarked it vs the one in Base? (esp. with respect to #8952)

@timholy I think the implementation in StatsBase has yet to be optimized. What about moving your PR there?

I'll give that a whirl (after a couple of other priorities).

I have thought about this, and generally have mixed feelings about not being able to do histograms in base julia. It is common enough that one may not want to install StatsBase and use it from a module.

I am ok with renaming, better APIs, but perhaps we should move Histogram from StatsBase to Base than the other way around.

Is it just me or does anyone else find the Matlab-style hist function really hard to use? If I do hist(v) where v is a vector of integers, it seems pretty obvious that I would want the values to be the bins. If we're going to have this functionality in base, I would want it to be better, more generic and faster.

Does one use hist much other than for plotting?

I tried to use hist to count the number of occurrences of each unique value in a vector, but I did not succeed. Is there a different function for doing that, or can it be included in hist?

We call that countmap in StatsBase.

Coming from a statistics background, I'm not really excited about mixing countmap and hist together. Histograms are designed for density estimation. Counting unique items may be related to density estimation (especially for finite sets), but I think of them as conceptually distinct tools.

Matlab seems to have deprecated hist and histc with histogram and histcounts. Not that it matters here in any way, but histogram draws, whereas you can get edges and such with histcounts. I don't think one needs a separate command for plotting histograms.

On this issue, I agree with @johnmyleswhite and consider a histogram a density estimate and what @StefanKarpinski asks for a frequency count or countmap (I think the latter word is more common in machine learning).

I also think of a histogram as the graphical representation and have been surprised by the hist in base. Therefore, I wonder if there is a separate interpretation of histograms outside classical statistics. Yesterday we got this question on the stats list

https://groups.google.com/forum/#!topic/julia-stats/CO3Pgc89Y7A

which was a use I was not aware of. Is it also used like this in CS?

Counting the occurrences of discrete items and summarizing the distribution of continuous values using bins (i.e. constructing histograms) are different concepts and usually used in different contexts. Obviously, we should use different functions for them.

We should remove this from Base in 0.4 in favour of the functionality in StatsBase.

Bump. A couple of the tests rely on hist functionality, should they be rewritten?

We can comment them out for now, and rewrite them, or move those tests to StatsBase.

+1 for removing, and also the older comments that it invites confusion (I tried to use it in the count-uniques sense earlier today...). StatsBase is great and very widely used - most users will have it installed anyway.

I'm fine with moving histogram functionality to StatsBase, but do have some issues with the implementation in StatsBase.

1. There does not seem to be an function to make a N dimensional histogram from a Array{T,N}. Instead one must input a NTuple{N,AbstractVector}. This seems strange to me, as any time I would want to compute a N-dimentional histogram the data would be in an Array. Also, the edges need to be input as a NTuple{N,AbstractVector} instead of a Array{T,N}, which seems awkward. Maybe I just didn't look in the right place?
2. Why is the function called "fit"? A histogram is computed directly from the input data and the edges. There is no fitting involved. I would much prefer to call the function hist or histogram as is more common and much more discoverable.

Having a function called hist is fine. But a histogram is both a descriptive statistic and an inferential statistic, so it's certainly a type of model that one fits to the data.

I've updated the PR.

@BobPortmann Those are good points: would you mind opening an issue on StatsBase?

Bump.

This was a 0.4.x milestone PR; still relevant? Should we remove hist & friends before 0.5 release?

Marking 0.5.0 for the triage team.

Closed in favour of #16450