My questions are related to multivariate time series classification, hence it may differ from forecasting problems. I can have either variable (entire history of the series) or fixed time steps (window of every 5 or 10 time steps) and about 50-60 features.
For CNN, does the order of the features matter? If the convolution layer learns from neighboring features, why would this not be the case?
Do LSTM and CNN benefit or harmed by features engineered to help traditional machine learning or tree-based algorithms. For example, if I want one of my classes to be defined as having
feature_1to be larger than
feature_2, should I engineer a new column with boolean values for
feature_1 > feature_2? Similarly, I have linear regression lines fitted to certain features. What is important is the significance of the fitting, i.e. the p values where anything less than 0.05 or 0.001 is considered a part of a class. Do I need to create a boolean column for
linear_feature_1 < 0.05? The p value can be extremely small if they are significant, so scaling should help in standardizing the values across all feature. Can these algorithms also learn that one of the important features is the signed difference between
feature_3and a constant
feature_4? Do I need to explicitly engineer the difference in a column
feature_3 - feature_4?
What will happen to a model where sometimes a feature is important and some other times it is irrelevant or just noise?