Nowadays, I have a matrix of high dimensional about genetic expression with 655 variables. I want to find the influentional variables. Although there are some methods such as logistic regression and PCA, however, I think most of them are model-based and cannot perform well in different situations. I just find engineers in computer science use Convolutional Neural Network (CNN) to analyze the picture and recognize the handwritting number. They compress the high amount of pixels into a small number of pixels by pooling and operation of convolution. So I wonder whether it is effective to apply the same way in genetic data? Or how to make a change to apply the thought of CNN into high dimensional data analysis? I am looking forward for your answer! Thank you!
Using the measurement method such as mean impact value (MIV) may be more suitable for the variable selection. The manifold algorithm should have the intrinsic dimension estimation first, and the variable with different background cannot be selected using the PCA directly. Statistics or the outlined measurement method can investigate the short-term or long-term mean impact value for the specified CNN. Various methods can be used for variable selection. Good luck!
I think CNN's are exploiting geometry of imagery being shown. As gene expression data seems does not have any visual representation I don't think CNN's are a good choice.
You could try Deep Belief Networks (DBN) or at least Restricted Boltzmann Machines (RBM) - there are plenty of variations right now. But again as far as I know expression data are usually pretty scarce - meaning only hundreds of data points (vectors) with high dimensionality (as you mentioned 655). I would say for such dimensionality 10000 data points (vectors) is lowest boundary. Friend of mine had actually abandoned neural networks in favor of SVM's because of that.
So I think reducing dimensionality is highly welcome.
I agree with Andree that standard deep learning methodology is not going to be much help here unless you have very many instances. However, reducing dimensionality via a feature selection approach is also plagued by dramatic "over-fitting", i.e. your results will strongly depend on the instance set. You could try something like elastic nets, or combination of this with GSEA (from the broad institute). Heavily bagged approaches also provide some help here.
If you can always split into a training and a internal validation set. This may be hard if you don't have enough instances though. I have attached a presentation that shows an example using a similar data set (I am guessing here) from breast cancer gene sets (example 1 in the deck). Maybe this would help?
CNN will not help you in this specific problem,..as CNN works on patterns. U may try it on one column data but any type of results may expected. A good choice for you here may be t-SNE (https://lvdmaaten.github.io/tsne/). It is an award winning method for dimentionality reduction.
The CNN uses a property of physical phenomenon effectively, that is,visual scene has spatial translation symmetry that can be written well by convolution operation. The CNN includes this property in its network architecture, and is success to reduce the number of parameters. Thus, I guess the CNN might not be effective for your data analysis. I think it is better to apply some full connection type deep learning, such like RBM, DBN, and so on. Moreover, if you have some insight to your data to reduce the number of parameters, you should try to include the knowledge in your network architecture or learning rule.
I think instead of CNN (Because CNN is basically designed for image data) you need to take a look at deep auto encoders. It will reduce your data dimensions by selection only important variables and it can also learn different patterns in your data for regression and classification.
- Badges
- Science topic