Sparse array of string types

[orenbenkiki]

I tried to create a sparse vector of strings, and it failed due to zero(String) not being defined. I'm using the empty string "" as the zero value. Any chance of enhancing SparseArrays to support this?

[putianyi889]

In theory the empty string should be one(String)

julia> one(String)
""

[orenbenkiki]

Interesting...

I think the empty string still makes sense in sparse string vectors though.

[SobhanMP]

You can always add these at your own risk as I have not idea what they will break.

julia> Base.zero(::String) = ""
julia> Base.iszero(x::String) = length(x) == 0
julia> sparse(["123", "", "234"])
3-element SparseVector{String, Int64} with 2 stored entries:
  [1]  =  "123"
  [3]  =  "234"

IMHO, it might make sense to have an internal zero and iszero so that it's possible to overwrite them for sparse arrays only.

[orenbenkiki]

If I were writing an application, sure. But I'm writing a library, so changing Base behavior this way is probably not a good idea.

Having SparseArrays.zero and SparseArrays.iszero which by default forward to Base but allow overriding for additional types would be nice.

[SobhanMP]

I agree.

@ViralBShah @dkarrasch @LilithHafner Does it make sense to you?

[putianyi889]

I'd like to have an overlay structure similar to Jutho/SparseArrayKit.jl#13 . However SparseArrays.jl can't depend on FillArrays.jl and I don't know how to work around nicely.

[dkarrasch]

Having SparseArrays.zero and SparseArrays.iszero which by default forward to Base but allow overriding for additional types would be nice.

That sounds very sensible. It provides more flexibility at seemingly no cost. I wonder though what applications one has in mind, if one defines zero as something that is not the neutral element w.r.t. addition. Is it just about storing sparse rectangular data in the sense that "most" elements equal a fixed ("trivial") element? Or does some sort of algebra also come into play?

[orenbenkiki]

First, zero (whether for Base or specialized in SparseArrays) works based on the type of the array element. This makes it impractical for using it for the "most common value" of an array.

Second, SparseArrays code assumes that the zero value is zero w.r.t addition and multiplication, and breaking this assumptions would require modifying it in many places - not to mention, breaking client code which also makes this assumption.

Having CompressedArrays working similarly to SparseArrays and allow for an arbitrary "most common" value per array instance would have been awesome (and would have reduced the need for things like nullable/masked arrays). However, somehow this idea never caught on...

[ViralBShah]

The problem is of abstractions. The SparseMatrix data type and its numeric types are designed for numerical sparse matrix linear algebra calculations. So extending to non-numeric types is always going to feel unnatural. It would be nice to get as much as we can get for free within this design - I do get that.

For strings, isn't it better to use a different data structure like a dictionary?

[LilithHafner]

There is a proposal floating around somewhere to support sparse arrays with custom "zero" values. Implementing that (either here or in a different package) would make a sparse array of strings fit naturally into the abstraction.

[LilithHafner]

I don't like this

SparseArrays.zero(::Type{String}) == ""

because SparseArrays assumes zero * x == zero, and so a sparse array of strings will behave oddly under concatenation. A structural empty string concatenated with anything result in the empty string while stored empty strings concatenated with anything (x) will result in that thing (x).

[orenbenkiki]

@ViralBShah:

It would be nice to get as much as we can get for free within this design - I do get that.

That's the idea.

For strings, isn't it better to use a different data structure like a dictionary?

Depends on the use case (of course). In my case, definitely not. Vectors have properties like length, and can be sliced, and masked, and so on. NamedArrays makes them dictionary-like, if that's needed.

@LilithHafner:

SparseArrays assumes zero * x == zero, and so a sparse array of strings will behave oddly under concatenation.

(As an aside, I would have been happier if string concatenation was considered to be addition rather than multiplication, commutativity be damned. E.g. sum(strings) would have worked for concatenating all the strings in a vector which makes sense. But that's just me.)

In what context does the above matter? Note that you can't add strings so trying to do linear algebra on strings would fail anyway.

There is a proposal floating around somewhere to support sparse arrays with custom "zero" values. Implementing that (either here or in a different package) would make a sparse array of strings fit naturally into the abstraction.

That would be nice.