Skip to content

Clustering

The Cluster tab groups the colonies of the Test Set by their features, with no model: useful to see what a plate holds, or to make classes from the data themselves.

The Cluster tab

Clustering replaces the model

GO throws away the current classes and any trained SVM, and puts the clusters in their place. If the model has not been saved, it first asks The Model is not saved. Continue?: Cancel stops, Yes and No both go on.

Running it

  1. Put the plates in the Test Set, and tick their colonies.
  2. Choose the filters and features on the Features tab. Clustering uses them directly: as the tab says, Cluster Analysis does not use the Feature Reduction.
  3. Choose an Algorithm, and press GO.

Baclan calculates the features, draws a histogram of the distances between colonies on the chart, and finds the clusters. Each colony's Class in the table becomes its cluster, the plate image colours its circle, and the legend counts them.

The classes of the Training Set are now Cluster_1, Cluster_2, ..., each holding copies of its colonies, plate by plate. You can look through them, untick mistakes, rename them, and train on them.

Algorithms

Algorithm Does Settings used
None Nothing
K-Means Splits the colonies into the number of clusters you ask for. It uses Euclidean distance, whatever Distance says Desired Number of Clusters, Number Of Restarts
Hierarchical Joins the closest colonies and clusters, step by step. You choose where to stop Distance, Cluster distance
R (ToDo) An experimental method that decides the number of clusters by itself, from the distances between colonies Distance
External Leaves the clustering to another program, below

Hierarchical clustering

A window shows the clusters at one step of the joining, one row of thumbnails per cluster. Choose the step with the box at the bottom, and the size of the thumbnails with the scroll bar. ShowD shows the distances. Ok cuts the tree at that step; Cancel at the step Baclan suggested.

Setting Choices
Distance, between two colonies Euclidean, Euclidean Relative, Earth Mover, Kolmogorov-Smirnov, KullBack-Leibler
Cluster distance, between two clusters Centroid, Min, Max, Mean, Ward

External clustering

  1. Baclan asks where to save the feature matrix, as a comma-separated file with one row per colony, and shows the External Cluster window.
  2. Cluster the file with your own program. Write the labels to a text file, one integer per line, in the same order, numbered from 0.
  3. Press Done and choose that file.