I am trying to cluster 3d objects based on their images and according to their level of detail. The main struggle i face is clustering with this high number of dimenssions. I tried the use of PCA to reduce the number of dimenssions, but it resulted in either loosing some details or in also high number of dimenssions. Any recomendations on:
1- Choice of algorithm: Kmeans is acting very poorly.
2-Distance metrics: In light of typr of data and algorithm.
3-Additional steps to reduce number offeatures.
4- Number of dimenssions: or an indication for acceptable number of features.
As a choice of model selection for multivariate data, I would like to suggest you multivariate statistical models from directional statistics or spherical statistics. The most popular such models are von Misses Fisher distribution, Kent distribution, Angular Central Gaussian distribution, Fisher-Bingham distribution on the special manifolds or the most effective multivariate Bingham antipodal symmetric distribution. Some related literature for you may be here: https://www.jmlr.org/papers/volume6/banerjee05a/banerjee05a.pdfhttps://arxiv.org/pdf/2005.06889.pdf
http://webspace.ship.edu/pgmarr/Geo441/Lectures/Lec%2016%20-%20Directional%20Statistics.pdf
Have you tried Self-Organizing Maps (SOM) ?
https://towardsdatascience.com/self-organizing-maps-1b7d2a84e065
These might help :
https://www.youtube.com/watch?v=sIyMAzAHw6I&t=467s
https://www.youtube.com/watch?v=0IjQ0SyQEOI
https://www.youtube.com/watch?v=MzmjQ3qI_kM
https://www.youtube.com/watch?v=ahNxYK8oEwE&t=2s
you can try to represent this data with a probabilty distribition function (PDF)
with algorithms such as Gaussian Mixture Model (GMM).
these methods represent a vast data through the parameters of an approximated PDF. there are also other PDFs such as HDP (Hierarchical Dirichlet Proccess)
the trick is to find the distribution which represent your data distribution in ther most accurate way.
You can also try tensors decompositions such as Parafac, Tucker to reduce the model and extract the features,
more details you can find in:Article Tensor decompositions for feature extraction and classificat...
orArticle Tensor Decompositions and Applications
- Badges
- Science topic
- Similar topics
- Instrumentation Engineering
- Instruments