I'm a theoretical physicist in need of some advice about creating a SVM for a data-science-esque job I managed to land somehow. I've never done any machine learning prior to this. I don't frequent this board, and I hope this is not close enough to a HW thread to get me banned.
I have a random vector X \in R^n which I need to classify into two groups, labelled 1 and -1. I've heuristically observed that X \in -1 generally have either high or low values in certain components of X, whilst X \in 1 generally fall "somewhere in-between". So far I have mostly done tests with "large" -1 data, because my dumb-ass disregarded the small data. It's very advantageous for my implementation to use a linear kernel.
My question is: due to the nature of X \in -1 and my desire to keep a linear kernel, should I add in another label? Or will the linear kernel be able to distinguish 1 from -1 given enough samples (R^2 intuition suggests no)? My data -1 could perhaps be manually labelled into two subsets as they are visually distinct, although this would end up being quite painful. One could probably train an auxiliary SVM to divide the data into small and medium-to-large and then clean up the mistakes manually, but this seems like poor form somehow. As the model needs to be maintainable, this could also be problematic in the future.
I am grateful for any advice, and could perhaps be convinced to swap kernels, but I'd rather make the linear kernel work. n \in [8,16] if that matters.
I have a random vector X \in R^n which I need to classify into two groups, labelled 1 and -1. I've heuristically observed that X \in -1 generally have either high or low values in certain components of X, whilst X \in 1 generally fall "somewhere in-between". So far I have mostly done tests with "large" -1 data, because my dumb-ass disregarded the small data. It's very advantageous for my implementation to use a linear kernel.
My question is: due to the nature of X \in -1 and my desire to keep a linear kernel, should I add in another label? Or will the linear kernel be able to distinguish 1 from -1 given enough samples (R^2 intuition suggests no)? My data -1 could perhaps be manually labelled into two subsets as they are visually distinct, although this would end up being quite painful. One could probably train an auxiliary SVM to divide the data into small and medium-to-large and then clean up the mistakes manually, but this seems like poor form somehow. As the model needs to be maintainable, this could also be problematic in the future.
I am grateful for any advice, and could perhaps be convinced to swap kernels, but I'd rather make the linear kernel work. n \in [8,16] if that matters.
