I have a two dimensional data, i.e. 2 variables (e.g. x and y) with 1500 observations each. Visual inspection shows clearly that the data are clustered according to a third dimension in a binary matter. Meaning that: for z threshold data are clustered.
I am looking for a clustering algorithm that can choose the best threshold that for example minimizes the variance. I need something so not to "just choose a threshold by eye".
See examples below:
Charalambos,
If you're looking for algorithm that can automatically cluster the data, you could try "Expectation Maximisation" clusterer. This algorithm assigns a probability distribution to each instance which indicates the probability of it belonging to each of the clusters. In addition, it decides the number of clusters using cross-validation method.
On the other hand, K-means algorithm tries to minimize the within cluster sum of squares (or variance), but the number of the generated clusters can be selected previously by the user.
HTH.
Samer
Thank you guys for the suggestion. I think that linear discriminative analysis (LDA) might be the best choice though, because it chooses the biggest distance with the smallest variances in the clusters. Bayesia and "Expectation Maximisation" seem to not have physical interpretation for my case because I have not any Probability distribution function (PDF) or any prior model. C4.5 seems to be a very good tool. It was also voted #1 machine learning tool in 2008, but it seems to complicated for the moment. I will keep looki.
Data science rules,
Cheers!
Dear Harry,
State clustering a data subset for a limited participation of space that you increase gradually. As you increase the Region of data there will be a size for which clustering becomes loose. Then you determine the Region where clustering is important. Repeat in other Regions.
Dear Professor,
I did that by eye, but I was looking a more "mathematical/statistical measure". Though, it not so important to separate the clusters with high precision, working on the corresponding dataset is a good practice for machine learning algorithms.
Charalambos,
Just wanted to highlight one point. In your question you said:"How can I cluster 2-D data using only a 3rd dimension?". This means that you are looking for a clusterer algorithm to be used, but later you referred to "linear discriminant analysis (LDA)" and C4.5, where both are *classification* algorithms (not cluster methods). Therefore, if you are looking for something similar to LDA, you may need to consider something like "logistic regression".
Cheers,
Samer
Samer
the current dataset could also be classified into two classes assuming that those have physical interpretation. For clustering, I tried kmeans and it did not succeed. Clustering methods might not be useful because the dataset spans many orders of magnitude and follows a power law trends. I might also check the logistic regression. Thanks for the recommendation.
Charalambos
We can transfer clusters to identical groups using anew algorithm then we don’t need to work on dimensions
- Badges
- Science topic
- Similar topics
- Computer Science
- Data Mining