I used rapidminer to preform clustering for column using kmeans. I want to know whether clustering preformed well or no , I used cluster preformance by distance (avg within centroid and Davies Bouldin)
cluster_0 centroid:-0.068
cluster_1 centroid: -1.354
cluster_2 centroid:1.129
PerformanceVector: Avg. within centroid distance: -0.218 Avg. within centroid distance_cluster_0: -0.120 Avg. within centroid distance_cluster_1: -0.309 Avg. within centroid distance_cluster_2: -0.313 Davies Bouldin: -0.591
i couldn't understand these results
Hi Hala Alrobassy
You choose Euclidean measurement as distance metric. So you measure the Sum Squared Error (SSE) between the centroids and the samples. And the minimum SSE value gives you the best centroids.
Best regards.
Silhouette Coefficient is an unsupervised metric ranging from -1 to 1 for measuring which clusters are well-formed and well-separated. So if your objects in clusters give a score close to 1 it means that your clustering is well-formed. While calculating silhouette coefficient you can apply to a distance metric which is suitable to your data (e.g. Manhattan or Euclidian Distance). Following sources might be helpful:
1. You can check an application of it in the following paper: Layton, Robert, Paul Watters, and Richard Dazeley. "Automated unsupervised authorship analysis using evidence accumulation clustering." Natural Language Engineering 19.1 (2013): 95-120.
2. If you are a python / scikit-learn user, you can check following links:
https://scikit-learn.org/stable/auto_examples/cluster/plot_kmeans_silhouette_analysis.html
https://scikit-learn.org/stable/modules/generated/sklearn.metrics.silhouette_score.html
3. If you need a MATLAB example: https://www.mathworks.com/help/stats/clustering.evaluation.silhouetteevaluation-class.html
so, why this results and Davies Bouldin appeared in negative ? i made standard normalization for the clustered column. I didn't use Silhouette Coefficient metric
Hala Alrobassy, in theory Davies Bouldin Index returns a value in between 0-1. However, according to the tool that you used or the parameters you can configured, index range might change. So, there is a similar discussion about it, which might be helpful to solve your problem: https://community.rapidminer.com/discussion/21631/davies-bouldin-index-returns-negative-values
Another point is, as @Murat says, your distance measurement would give a positive value. So you can also check if you are using a proper distance measurement method or not.
What do you mean with standard normalization: zero mean unit variance? If you perform this normalization on distance results, it would change the output of Bouldin Index. Your can try shifting your normalized distances to positive direction in order to have non zero values.
Dear Hala Alrobassy
You can select the "maximize" option to get positive values for Davies Bouldin.
The near value to zero in both cases is better.
Sincerely
- Badges
- Science topic
- Similar topics
- Computer Science
- Computing