added t-sne docs & tests

Author: wildart

docs/make.jl

@@ -14,6 +14,7 @@ makedocs(
 		 	"Laplacian Eigenmaps" => "lem.md",
 		 	"Local Tangent Space Alignment" => "ltsa.md",
 			"Diffusion maps" => "diffmap.md",
+			"t-SNE" => "tsne.md",
 		],
 		"Misc" => [
 			"Interface" => "interface.md",

docs/src/index.md

@@ -43,5 +43,6 @@ Following dimensionality reduction methods are implemented in this package:
 |[`LEM`](@ref)| Laplacian Eigenmaps |
 |[`LTSA`](@ref)| Local Tangent Space Alignment |
 |[`DiffMap`](@ref)| Diffusion maps |
+|[`TSNE`](@ref)| t-Distributed Stochastic Neighborhood Embedding |
 
 **Notes:** All methods implemented in this package adopt the column-major convention of JuliaStats: in a data matrix, each column corresponds to a sample/observation, while each row corresponds to a feature (variable or attribute).

docs/src/tsne.md

@@ -0,0 +1,22 @@
+# t-Distributed Stochastic Neighborhood Embedding
+
+The [`t`-Distributed Stochastic Neighborhood Embedding (t-SNE)](https://en.wikipedia.org/wiki/T-distributed_stochastic_neighbor_embedding) is a statistical dimensionality reduction
+methods, based on the original SNE[^1] method with t-distributed variant[^2].
+The method constructs a probability distribution over pairwise distances in
+the data original space, and then optimizes a similar probability distribution of
+the pairwise distances of low-dimensional embedding of the data by minimizing 
+the [Kullback-Leibler divergence](https://en.wikipedia.org/wiki/Kullback%E2%80%93Leibler_divergence) between two distributions.
+
+This package defines a [`TSNE`](@ref) type to represent a t-SNE model, and provides
+a set of methods to access its properties.
+
+```@docs
+TSNE
+fit(::Type{TSNE}, X::AbstractArray{T,2}) where {T<:Real}
+predict(R::TSNE)
+```
+
+# References
+[^1]: Hinton, G. E., & Roweis, S. (2002). Stochastic neighbor embedding. Advances in neural information processing systems, 15.
+[^2]: Van der Maaten, L., & Hinton, G. (2008). Visualizing data using t-SNE. Journal of machine learning research, 9(11).
+

test/runtests.jl

@@ -87,14 +87,17 @@ end
     X, L = ManifoldLearning.swiss_roll(n; rng=rng)
 
     # test algorithms
-    @testset for algorithm in [Isomap, LEM, LLE, HLLE, LTSA, DiffMap]
+    @testset for algorithm in [Isomap, LEM, LLE, HLLE, LTSA, DiffMap, TSNE]
+        #print("$algorithm ")
         for (k, T) in zip([5, 12], [Float32, Float64])
             X = convert(Matrix{T}, X)
 
             # construct KW parameters
             kwargs = [:maxoutdim=>d]
-            if algorithm == DiffMap
+            if algorithm === DiffMap
                 push!(kwargs, :t => k)
+            elseif algorithm === TSNE
+                push!(kwargs, :p => k)
             else
                 push!(kwargs, :k => k)
             end
@@ -114,13 +117,15 @@ end
             @test eltype(Y) === T
             @test size(M) == (3, d)
             @test length(split(sprint(show, M), '\n')) > 1
-            @test length(eigvals(M)) == d
 
             # additional options
-            if algorithm !== DiffMap
+            if algorithm !== DiffMap && algorithm !== TSNE
                 @test neighbors(M) == k
                 @test length(vertices(M)) > 1
             end
+            if algorithm !== TSNE
+                @test length(eigvals(M)) == d
+            end
             if algorithm === LEM
                 @testset for L in [:sym, :rw]
                     Y = fit(algorithm, X; laplacian=L, kwargs...) |> predict