Feature extraction and classification of signal components belong to very common problems in various engineering, economical and biomedical applications. In this question i wish to know whether these feature extraction techniques hold good for almost all application or it varies with nature of the problem and data.
Without a doubt, the feature extraction methods can be problem specific. One approach is to throw all the features you have in to a feature selection algorithm (see the wikipedia link) and then provide these optimal feature sets to a classifier. If you have no clue what features to use, and you have a lot of training data, you can try a neural network approach and achieve good results on the data directly. But often an insight in to the characteristics of your problem will guide you to an approach feature set. For example, you have signals of interest that are short and intense with a scale behaviour, so you consider wavelets. Or a collection of fundamental tones of longer duration characterizes your signal, so you try a Fourier basis. Or in 2D, you can see that your signals of interest have fine curve structures so you start with a curvelet transform. Or a more constrained training approach using data dictionaries.
The variations are immense, but the best results will be obtained by undertaking a period of "data exploration" upfront where you try to gain some insight in to the characteristics of the problem you have. Then you work from there.
https://en.wikipedia.org/wiki/Feature_selection
when your input data in any algorithm is too large in values or is going to be redudent then we can transform the data in a specific form which is easy to use and handle. Different types of transforms provides us the methods to transform our data.
- Badges
- Science topic
- Similar topics
- Analytical Chemistry
- Extraction