Potential cross-subject windowing issue in `__getsamples()` for fMRI data

[wenruifan]

Hi, thank you for sharing this repository.

I noticed that the provided demo fMRI data are shaped as [N_subjects*N_timesteps, N_nodes].

I have a question regarding the sampling strategy used in Utils/Data_utils/real_datasets.py, specifically the __getsamples() function.

def __getsamples(self, data, proportion, seed):
    x = np.zeros((self.sample_num_total, self.window, self.var_num))
    for i in range(self.sample_num_total):
        start = i
        end = i + self.window
        x[i, :, :] = data[start:end, :]

In this function, sliding windows are constructed with a stride of 1 over the entire N_subjects*N_timesteps dimension of the input data.

For fMRI data, if the input data are formed by concatenating multiple subjects or sessions along the temporal axis, this approach may generate windows that span across subject or session boundaries, for example containing the last time points of one subject followed by the first time points of another.

I would like to clarify the intended usage:

If multi-subject data are concatenated, is it intended that windows may cross subject or session boundaries, effectively treating the data as one continuous time series?

Clarifying this assumption would be very helpful for correctly applying the method to multi-subject fMRI datasets. Thank you very much for your time and clarification.