Sammon mapping software

While PCA simply maximizes variance, we now want instead to maximize some other measure, that represents the degree to which complex structure is preserved by the transformation.

Various such measures exist, and one of these defines the so-called Sammon Mapping, named after John Sammon, Jr, who initially proposed it. PCA has a drawback. If all the data variables have different variances then the contribution of each of these variables towards the output will be much lesser. Hence, scaling is required to bring all the variables to a similar variance, between 0 and 1. And then the contribution to all variables can be calculated. Unlike traditional linear dimensionality reduction techniques such as PCA , the Sammon mapping does not explicitly represent the transformation function.

Instead, it simply provides a measure of how well the result of a transformation i. We can easily picture data in 2D as well as 3D but anything higher than these dimensions, it is difficult for us to visualize. In such a case, we need to project whole datasets to a lower dimension i. Some linear projection algorithms are used frequently for such cases. We all have heard about PCA Principal Component Analysis , if not then, PCA follows static orthogonal projection to get linearly uncorrelated points in lesser dimension.

Although PCA maximizes the original variance present in the transformed dataset, yet it shows problem while projecting some structured regular pattern in a curved manifold. For example in figure 1, I tried to plot a bouquet of circles in 6d, each circle perpendicular to each other using PCA.

As a result, we get:. Fig 1. Projection of Bouquet of Circles passing through the origin, where each circle is perpendicular to each other, using PCA. So, Sammon mapping tries to eliminate such limitation. What it does is that it takes the result of PCA with maximum variance and adds other features what feature? Subsequently, we have to remember a point that Sammon mapping does not give the complete solution but it rather tries to minimize the error and provides the optimal result.

It tries to minimize the difference between the distance of data-points with each other in transformed space and original space.

You signed in with another tab or window. Reload to refresh your session. You signed out in another tab or window. On completion, y contains the corresponding. By default, a two-dimensional. A set of optimisation options can be specified using optional. The default options are retrieved by calling sammon x with no. File : sammon. Date : 18 April Authors : Tom J. Sammon mapping in Python 23 stars 17 forks. Branches Tags.

Could not load branches. Could not load tags. Latest commit. Git stats 22 commits. Failed to load latest commit information. View code.