Change the code so that whatever dataset can be plugged in. Now it must be a TF module.

Try to make the execution faster on datasets when possible (e.g., use vectorized functions or apply_raw)

• Remove the heatmaps overlay;
• Remove the ticks and values from the axis but keep the frame;
• Change the tiles from the predicted label to an ordered label (i.e., A B C D);
• Export 4 clusters with all the possible combinations;

Correlate the agreement measured during the understandability pilot study with the purity of the cluster.

Correlate the features chosen by the evaluators during the sufficiency pilot study for the high-level approaches with the features' combinations actually used.

Correlate the regions chosen by the evaluators during the sufficiency pilot study for the low-level approaches with the ranking of the regions.

The images for the labelability should show two taggers at a time, the RQ is whether it exists a property such that it distinguishes the two clusters.
The clusters are chosen based on the number of misclassified elements in the cluster.
When choosing the images, we consider only images of misclassified entries. We select n images as the medoid of the cluster plus its n - 1 closest neighbors.

Make sure to repeat all the non-deterministic steps in all the runs.

Add the information about the perplexity value for the low-level approaches in the merged_data dataset. The information can be used when plotting the images to show the complete name of the low-level approaches.