Plot Classifier Confusions

 

After training, we can look at the confusion matrix of the predictions between the two models.

  1. Select your classifier from the classification pane.
  2. Click on Plot Confusions in the classification panel.
    • Then you can decide if you want to exclude any images from the confusion matrix.

    Classification_PlotConfusions_Exclude

  3. Click OK and the confusion matrix will be generated:

    Classification_PlotConfusions

  4. You can export the confusion matrix as a .pdf or .png.
  • The predictions for model 1 are on the y-axis and predictions for model 2 are on the x-axis.
  • The column of percentages on the right side of the plot shows for the cells that model 1 predicted as cell type X the fraction where model 2 also said X. For example, 80% of cells that model 1 predicted to be Bcell were also predicted to be Bcell by model 2.
  • The row of percentages at the bottom is the same but for model 2. So 59% of cells that model 2 predicted to be Bcell were also predicted to be Bcell by model 1.
  • Do not be startled by low numbers in the beginning, as you train cycles the two models will learn and converge.
  • This plot can give you an idea of the most difficult cell types to classify or the most frequent confusions.
  • By default CellTune takes these values into account when sampling. It tries to balance agreement across cell types by sampling more from disagreements where one of the predictions is of a cell type with low agreement.