The K-means clustering algorithm is sensitive to outliers, because a mean is easily influenced by extreme values. K-medoids clustering is a variant of K-means that is more robust to noises and outliers. Instead of using the mean point as the center of a cluster, K-medoids uses an actual point in the cluster to represent it. Medoid is the most centrally located object of the cluster, with minimum sum of distances to other points.
From the above pictures, we can see that the algorithm correctly arranged the pictures into clusters. Algorithm arranged images into clusters based on medoids, which measure the distances between images or vectors by the distance based on cosine similarity. This algorithm divided the images into clusters according to their type, i.e. there are only flowers of the same species in each cluster. In the first cluster, there are daisies, in the second dandelion, in the third roses, in the fourth sunflowers, and in the last, fifth, are tulips.
I think the results are very good, but that doesn't mean that our algorithm works perfectly on real-world problems. Opposite! We probably got such good results because of the selected pictures, which are very similar to each other within the species, and the species as a whole differ greatly from each other. Similar results in the given images would likely be obtained with an algorithm that does not use a neural network. Due to the mentioned opinion, we tested our algorithm in other image domains, but where it was not cut off so well. For example, pictures of vehicles (car, bus, suv, and tractor) were grouped only partially correct, and in the pictures of food (hamburger, pizza, pasta, salad) the resulting lumps were completely wrong.
To sum up, the algorithm does not always group images correctly according to content, which means that we can help classify the data into approximate, real-world problems with our algorithm clusters, but the accuracy of these clusters cannot be completely relied upon.
The number of groups was determined based on a priori knowledge of the problem domain, since the images above which we made the findings, we chose ourselves. We chose exactly five different types of flowers, namely daisies (cluster 1), dandelion (cluster 2), roses (cluster 3), sunflowers (cluster 4), and tulips (cluster 5).
The algorithm divided the given images into five clusters with an average silhouette of 0.397. To interesting findings, we came up with trying all the other numbers groups. For example, we obtained an average silhouette for three clusters 0.412, for four clusters 0.381, for six clusters 0.403 and for seven clusters 0.385.