I'm taking my first steps in data mining while working on a student project. I'd appreciate any leads on the following:
I have a data set with the following properties:
- a set of Features columns (let's say, 10);
- a set of Time Series columns (let's say, 7). Please note, that time representation is sequential (e.g. day 1, day 2, ..., day 7);
- a number of rows is high (let's say, 1k).
- some feature values, as well as some time series values may be missing (blanks and NaNs)
- the correlation between rows is very high (for example, row #16 and 17 have coef. of cor. ~1)
- the correlation between individual features and time series is very low (~0)
My objective is to use both sets to predict time series column #8.
The approach I have been thinking about goes something like this:
Since the same features represent different values for the time series (in one row there are 10 values for features which are the same for the 7 values of time series data) i'm looking for a method to extract a matrix? of 7 values and 1k rows that will play a role of "unique identifier" for each value of time series in a row (So that we will be able to calculate the value, for instance, of the time series column 5 using features and that missing matrix). Later on, I plan to fill in the gaps of missing feature values and time series and later apply standard prediction analysis methods.
Based on my experience I want to look into more details on Granger Causality analysis and nonlinear autoregressive exogenous model (NARX) which is a nonlinear autoregressive model.
Any hints would be kindly appreciated, including if you are able to tell me what this sort of problem is called so that I can research further.