For dimensionality reduction we use PCA , and for clustering we use K means. How we can combine them simultaneously?
You can also consider K-means as a dimension reduction technique. If you have an N dimension data and you cluster them into C clusters, then you basically reduced the dimension from N to C. If you are using a a hard K-means clustering then for sure the output will be binary and not useful for dimension reduction. My suggestion is to use K-means first and then use the cluster centroids as a basis for Radial basis functions. You can check the attached paper, especially section 5.3.
Article Soft Radial Basis Cellular Neural Network (SRB-CNN) based Ro...
.
spectral clustering might be what you are looking for ; the following "quick overview" does a great job at describing what it does and gives the core references :
https://charlesmartin14.wordpress.com/2012/10/09/spectral-clustering/
.
Hussain,
It's a very sensitive procedure, and we apply it as follows:
We build a filter that allows the training set to pass through PCA filter first, and then it gets clustered by the base clusterer method (e.g., K-means), while the test set remains untouched.
HTH.
Samer
Dear Hussain Sargana
I have no idea regarding your data? if it is time series data it must be prepossessed properly because field of feature engineering have gained prominent attention where traditional techniques have become unavailable to support.
I would recommend you to see features of your data then decide what to do?
Best Wishes,.
JAMIL AHMED
Establish the features of your dataset. Build a vector of features and then use an algorithm to reduce the dimension of the vector (the simplest one is 'greedy algorithm')
Dear Hussain,
Not sure about dimensions of your data. Hope this links may help you.
http://www.sciencedirect.com/science/article/pii/S0957417414007301
ieeexplore.ieee.org/iel5/6012258/6019524/06019845.pdf
www.toknowpress.net/ISBN/978-961-6914-09-3/papers/ML14-521.pdf
Dear Hussain,
We recently published a new technique related to the selective clustering and dimension reduction. It may be interesting for you.
Conference Paper Weighted Spectral Cluster Ensemble
Clustering techniques can be used for dimensionality reduction problem also. But, it depends on the type of data also. So, similarity issue among the data is main concerned here. Based on which, we can choose hard clustering or soft clustering approach for this. K-Means is a famous one for the hard type of approach for this. A little modification of this algorithm will give you a better result. Otherwise, you have FCM for soft based approach. Distance metric selected plays a very important role in dimension reduction using the clustering technique. Following are the links for paper that may give you an idea about the clustering techniques and distance metric impact on them.
1>A comparative study between fuzzy clustering algorithm and hard clustering algorithm
http://arxiv.org/pdf/1404.6059
2>Effect of Different Distance Measures on the Performance of K-Means Algorithm: An Experimental Study in Matlab
http://arxiv.org/pdf/1405.7471
Thank-you all folks , your contribution is really really appreciable !
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Algorithms