Hi Frank, I am sorry I did not explain it clearly. I planned to give you an example to demonstrate the usage of GWR. But there are a number of problems with my connection and the demo was cancelled.
To answer your questions: 1) As for the required operations or functions, some common functions are involved in the process of estimation including extraction of the diagonal elements for a specified matrix to generate matrix or vector, and AIC/AICc and Adj-R2 function for linear regression. In addition, a high performance Weighted Least Square(WLS) fitting function ( e.g. , QR-based ) which GWR is mainly based on will be beneficial for implementation.I am not sure all these operation are available in current package. If these functions are not available, I will implement them. 2) The computational burden of GWR is loop of fitting on every regression points. The steps of a loop are independent because they only need to write coefficients for individual steps. GWR is similar with doing several OLS on every steps with individual weights. I think we can divide the entire loop into several group to parallelize the algorithm. I think we can utilize openmp on single multicore machine and MPP technology on GPDB. I will also welcome any suggestion you may have at further improving the functionality of parallelized implementation so that it can better serve our (future) needs. 3) Raster type is recently supported type in PostGIS. And the good news is that, as Iran mentioned, GPDB will support raster type in the near future. Maybe we could disable raster function when raster supporting is not available . 4) It is the recommended least sample size for GWR. According to P ez et al.(2011), the basic GWR is not an appropriate method for small sample sizes (<160). The larger the sample size the more accurate estimation we have. However, there is no existing GWR developed for mass data. Our colleagues are always going on about the poor performance and inefficiency of performing GWR on large amounts of data. Best, Chenliang Wang On 02/17/2016 10:21 AM, Frank McQuillan wrote: > Thank you to Chenliang Wang for presenting Geographically Weighted > Regression (GWR) analysis of spatial data at the MADlib community meeting. > > Here are some follow up questions that we did not get to in the meeting. > Chenliang, could you briefly respond? > > 1) what type of matrix operations are required? > > 2) how will the algorithm be parallelized? > > 3) is raster support for PostGIS a requirement? (GPDB currently does not > support as per > http://gpdb.docs.pivotal.io/4340/ref_guide/postGIS.html#topic_wy2_rkb_3p) > > 4) what does the 160 refer to in your slides? > > Thanks again, > Frank >
