Manifold Alignment

We have two slightly different versions of alignment approaches. The input to represent the weight matrix for each domain can be given in two forms:

1. a sparse matrix.
2. an array modeling the k-nearest neighbor for each instance. The files for this are in Alignment/knn/.

Folder: Alignment

Affine Matching

• AffineMatching.m

Procrustes Alignment

• Procrustes.m

Canonical Correlation Analysis (CCA)

• CCATwo.m
• CCAThree.m

Feature-level Manifold Projections

• wmapGeneralThree.m
• wmapGeneralTwo.m

Instance-level Manifold Projections

• wmapGeneralThreeInstance.m
• wmapGeneralTwoInstance.m

Unsupervised Alignment

The code is the same as feature-level manifold projections. The only difference is how to create the correspondence matrix.

• generateWeight3.m, used to generate weight matrix, calls:
  • computeOptimalMatch.m
  • decompose3.m

Folder: Utility

• cmpEmbedding.m - compares different embedding results. Examples are represented as columns.
• knnsearch.m - k-nearest neighbor search
• LaplacianEigenmaps.m
• LPP.m - Locality Preserving Projections
• Showtopics.m
• createAllConnectedGraph.m
• createKnnGraph.m
• L2_distance.m - All pairs Euclidean distance