make theiler window argument of `predict_timeseries` about timeseriesprediction.jl HOT 12 CLOSED

As far as I can tell, the current implementation of theiler windows does not work here.
The theiler neighborhoods require an index n for the query, which does not exist in general.

Instead we would need a method that makes sure the nearest neighbors are not temporal neighbors of eachother....

from timeseriesprediction.jl.

The theiler neighborhoods require an index n for the query, which does not exist in general.

Damn, you are right.

from timeseriesprediction.jl.

Well, on the other hand, you can still find the closest neighbor and then not allow temporal neighbors of the closest neighbor. See here: JuliaDynamics/DynamicalSystemsBase.jl#45

from timeseriesprediction.jl.

What do you think of this?
could something like this work?

function neighborhood_and_distances(point::AbstractVector,R::AbstractDataset, tree,
                      ntype::FixedMassNeighborhood; w::Int = 1)
    ind = Int[]
    function f(i)
        for j ∈ ind
            if abs(i-j) < w
                return true
            end
        end
        push!(ind,i)
        return false
    end
    idxs,dists = knn(tree, point, ntype.K, false, f)
    return idxs,dists
end

from timeseriesprediction.jl.

Nice approach. Seems like it would work, but may have performance issues.

If it is bad, you can increase the performance by making f a Functor instead, with the only field a Vector{Int}, I guess...

from timeseriesprediction.jl.

@JonasIsensee I think this is a nice approach. We should have it in DynamicalSystemsBase neighborhood as well.

from timeseriesprediction.jl.

@JonasIsensee the theiler window is currently not an argument of localmodel_tsp. Make it a keyword argument that defaults to mean(τ) for methods that have τ and to mean(R.delay) for methods that given an abstract dataset.

from timeseriesprediction.jl.

Then we can close this issue.

from timeseriesprediction.jl.

Which implementation of it do you want though?

from timeseriesprediction.jl.

The one we have already. That excludes temporal neighbors of the query point only.

The only thing that is left is to only have the argument to tsp a swell.

from timeseriesprediction.jl.

The theiler neighborhoods require an index n for the query, which does not exist in general.

Damn, you are right.

from timeseriesprediction.jl.

my brain is literally out of order.

from timeseriesprediction.jl.