I am using different clustering methods which gives different clustering output. I need to benchmark to find out which clustering method is the best. I have 20-25 clusters and 25000 samples. I observed that number of sample in each cluster is different from different clustering method. If I use Adjusted Rand Index (ARI) then it just compute the similarity between two output. But it does not tell precisely how much accuracy each cluster might have it. Could I compute these with any other measurement like silhouette or somet other way?
Thanks.
I think only an expert who is well versed in this subject area can evaluate which clustering results are better in each specific case.
https://www.researchgate.net/publication/359768613
There are several measures to evaluate clustering results like silhouette and CH index. You can averages the outputs of these different measures and decides the best clustering method for your application.
Silhouette score is another commonly used metric that measures how well a sample is assigned to its cluster compared to others. This metric can help you evaluate the quality of individual groups and the overall clustering performance.
If I understand correct then Silhouette score measure the convexity of the cluster. If the clusters are not convex then this score can mislead. Is there any score which can measure the membership similiarity between two clusters. I was thinking for each of the cluster calculate the [intersection/union] ratio and find the avg for the common clusters.
Clustering methods try to cluster based on methods (partition or hierarchical), similarity functions (which may use different property of data) or heuristics ( number of clusters). Depending on the data, even k-means when rerun on same data set, may give you different clusters. Some border data points may change their cluster.
So best way is to take domain expert help to understand the clusters which explain the data best.
In absence of that use of measures like ARI or SSE can guide. I would recommend to remove outliers from each cluster before using these measures.
To compare the clustering output from the different methods depending on the size of the dataset, number of clusters, type of dataset, and type of software used.
- Badges
- Science topic
- Similar topics
- Mathematics