mgwr

Code Structure

• mgwr.py includes the MGWR class.
  • scipy_optimize() is the leading function.
  • A few functions are unique for gradient-MGWR:
    • scale_coord(), scale_distance() scales the coordinates/distance (to unitless) before optimization. the scale is controlled by coords_scale_factor.
    • reverse_scaling_distance() is the opposite process, to provide meaning distance (with unit) after optimization.
    • dist_nb_smoothing() builds projection between number of neighbors and distance. the key idea is introduced in the paper. some of the relevant codes are in kernels.py.
  • Some functions do not have to be strictly followed:
    • this includes set_up(), call_back(), set_bound(), set_start(), careate_output_table(), summary_excel(), result_store().
  • Other functions are no longer used.
    • newton.py() does not use scipy module but the optimization performance is bad.
    • pysal_exec() calls pysal library. but MGWR in pysal is not parallel so we don't use it anymore.
  • This module can execute GWR, but scripts relating to GWR are not well organized and it is slow. we don't have to use those.
• gradients.py holds the core algorithm of gradient derivation.
  • gradient_mgwr(), loop_nk(), matrix_calc() should be developed.
  • ignore all other functions. gradient_gwr() is functioning, but it's slow.
• kernels.py holds the algorithm to compute kernels.
  • we need to develop this because they includes some gradient derivation.
  • we have four kernel types here (distance/neighbor + guassian/bisquare). part of gradient are derived here, based on the chain rule.
• ignore other scripts mgwr/gradients_func_ij.py, mgwr/gradients_multi.py, optimize.py.