Hi everyone,
I want to ask for some general advice regarding my input.
I want to use machine learning to quantify chemical concentrations.
I want to use multiple different techniques to collect my data, and input them into a machine learning algorithm. As such my label would just be the analyte concentration, and my features would be the data I collect from a range of chemical techniques on a sample of known analyte concentrations.
Say technique 1 gives me variable1(y) vs variable2(x) with say a size of 2x100. Then technique 2 gives me variable3(y) vs variable 4(x) with a size of 2x500. All 4 variables are unrelated.
Would simply dropping the independent variables (x) and concatenate the dependent variables (y) into a 1x600 input theoretically be reasonable given that I keep the x variables the same between all my samples?
In such a case, is the different size of data from the two techniques a problem? In a sense that since there is much more data from technique 2, will the neural network treat it as inherently more important, or the weights will decide the importance during training and so it is not an issue. If it is an issue I could probably get the data to be similarly sized by doing 1D convolutions and pooling separately for each technique before I concatenate them.
Lastly, say my technique third technique collects similar 2D data of variable 5 (y) vs variable 6 (x) but with an extra variable being adjusted for 6 different values (variable 7) , giving a 3D data of 2x200x6. Could I similarly just drop the independent variable and concatenate this to the end end as well, giving an extra 1200x1 input to the other two techniques? Given of course that the 6 variables are always the same and the dependent variable I dropped I keep the same between all samples?
Thanks for the help. I am just a bit confused as the time series forecasting examples have a shared x-variable of time, and the other structured data examples had a 1D series of features associated with the label. Literature examples tend to use just a single technique and have the input as its 2D or 3D output, but I did not find anything that combines multiple techniques as an input into machine learning.
Best regards,
Dominik