cwrowley / dmdtools Goto Github PK

Hello, I find your paper "Dynamic mode decomposition for large and streaming datasets" very helpful. Excellent work! Yet I have a small question about the paper and the code.

There are four steps in the streaming DMD. Is it better if we switch the order of step 3 and 4? Because if we do compression in step 3 first, in step 4 we will lose some information gained from the expansion in step 2. If we switch 3 and 4, we make full use of step 2 regardless of the compression.

Run sdmd_run.m, see the original(first figure) vs the switched(second figure). The switched version did improve the result.

I have data which has 50000 rows. In that 41 features and 23classes. Is it possible to apply dynamic mode decomposition. If so can you give an example for classifying my dataset

It would be nice to add kernel DMD as an option, for instance to standard DMD or streaming (or potentially TLS). One possibility would be something like:

sdmd = StreamingDMD(kernel=10)

for a kernel (1 + x'y)^{10}, or

sdmd = StreamingDMD(kernel=f)

for a user-defined kernel f(x,y). Any thoughts?