Reshaping data for regression

[adampointer]

Hi,

How would you go about reshaping data for a regression task? I am a bit confused about how I go from a 2d format, for example 1000 data points long with 2 channels to the 3d format. Would the end result be (1000, 2, 10) assuming I want 10 lags 'look back'? Are there any utility functions to help or any good examples?

[baraline]

Hi @adampointer, could you please specify the following so we can better answer your question :

• When you say "a regression task", are you talking about forecasting or extrinsic regression ? I would assume forecasting since you talk about "look back".
• Did you want to use a specific estimator? (We don't have much currently as we want to implement them ourselves and not wrap other libraries, but more are coming with GSoC !)
• What do you want to accomplish? Again i'm guessing forecast the next 10 values but prefer to be sure.
• If you have some example code of what you are currently trying to accomplish, it can be useful too.

Thanks !

[adampointer]

Hey, I was thinking specifically about extrinsic regression. I have a set of independent features and I want to model their relationship with the target variable, with the end goal of forecasting the target variable a few timesteps into the future. To do this I was going to shift the target variable. I was looking at the convolutional regressors. Maybe I am getting muddled up with regression and forecasting, but it looks like forecasting is limited to using the previous history of the target itself and future-knowable exogenous features? However, I want to forecast one time series but using a collection of other time series's, none of which are known in advance. Does this make sense?

[baraline]

To clear out some confusion about the task type :

• extrinsic regression is about predicting continuous targets y (an array of shape (n_samples)) from a collection of series X (an array of shape (n_samples, n_channels, n_timepoints). These targets are not related to the values the series will take in the future.
  A made up example would be "Given a series X representing caffeine level in a body of 24h, predict the quantity of coffee drank in milliliters during the 24h period".
• For forecasting in the case of collection estimators you would have X : (n_samples, n_channels, n_timepoints) and y : (n_samples, n_timepoints), but most forecasting algorithms work on a single series giving you X : (n_channels, n_timepoints) the exogenous variables and y : (n_timepoints) your target.
  A made up example (without exogenous) would be "Given a series y representing caffeine level in a body of 24h, predict the caffeine level in 3h".

You can use exogenous features in forecasting with the exog parameter, but while we have the interface set up, we didn't yet implement an estimator that supports exogenous features.

So If I understand correctly, you have one time series X of shape (n_channels, n_timepoints) being your exogenous variable and another series y of shape (n_timepoints) your target right ?

That part got me confused :

However, I want to forecast one time series but using a collection of other time series's, none of which are known in advance. Does this make sense?

What do you mean by a collection here ?

[adampointer]

Sorry I explained this terribly. I am talking about multivariate forecasting using the RegressionForecaster, I guess this is not supported yet.

[TonyBagnall]

hi, we do not have a tools to automatically format multivariate forecasting yet, forecasting module is very much in development, but its possible to rephrase it as regression and use a standard regressor. This formulation is different to ETS/ARMA style and it is a cause of confusion :) We will be doing a lot more with forecasting over the Summer.

Suppose your series to forecast is (1000,) and you want to use two exogenous variables stored in array of shape (2,1000). You want to use a window length 100. So if you have a horizon of 1, to forecast point 101, you use points 1 to 100 of the original series, and (2,100) of exogenous.

This makes a single case of shape (3,100) to predict one value of y.
A collection is then all of windows so created. You form a collection of shape X=(900,3,100), y=(900,).

You can now evaluate a regressor in this data,

from aeon.regression import DummyRegressor
d = DummyRegressor()
d.fit(X,y)
a = d.score(X,y)

But this is of course biased, as you are evaluating on the same data you fit. Simplest solution is then to split into train and test, making sure all test are after train, say trainX = (600,3,100), trainy = (600,) the first 600 points, test the other

from aeon.regression import DummyRegressor
d = DummyRegressor()
d.fit(trainX,trainy)
pred = d.predict(testy)
a = d.score(testX,testy)

there are other models and how to evaluate forecasting in this way is much debated. We can add exog into the RegressionForecaster in this way really easily using this model, I'll make an issue, but it is not the only way of regressing, we could have an unequal length window of "all data before the forecasting point" which makes a regression of unequal length problem.

Note I have used DummyRegressor, but we have lots of actually useful ones in the regression module, including deep learners

[adampointer]

This is exactly what I was after. Thanks. Are there any utility functions to help reshape the 2d sklearn style input for this? It's not too hard to do myself, but I don't want to reinvent the wheel.