Include Fergus semi-supervised label propagation?

[adrinjalali]

It seems like scikit-learn/scikit-learn#1185 would be a candidate to be included here. Closing that other PR, and we can decide if it can be included here.

From the original PR:

---

This is a new semi-supervised algorithm based on Fergus et al 2009 [1].

Strengths:

• Closed-form solution, no need for iterations or convergence threshold
• Similar accuracy as other sklearn semi-supervised algorithms with moons data, much better accuracy than other algorithms with concentric circles
• In addition to 'knn' and 'rbf' modes, has an 'img' mode for segmenting images

Weaknesses:

• Full order of magnitude slower than other sklearn algorithms, likely due to eigh() on the graph laplacian (somewhat mitigated by use of sparse data structures)
• Eigen-decomposition precludes full integration with existing semi-supervised class structure
• Method for computing image affinities (img_rbf_kernel) is inefficient and could use some optimization

I would love to add this algorithm to scikit-learn's semi-supervised package. I'm definitely open to suggestions for improving it and bringing it up to par with scikit-learn's coding standards and algorithm performance.

[1] Rob Fergus et al. Semi-supervised learning in gigantic image collections, 2009.