Skip to content

Tang2020_ASDFactors: Unexpected K=3 polarLDA runs with 0/NaN likelihood and very large alpha #63

Description

@wxinlei210-bit

Helllo,

I am using stable_projects/disorder_subtypes/Tang2020_ASDFactors with polarLDA on my own functional connectivity dataset. For K=3, I ran Step2A with 100 initializations, followed by Step2B.

I noticed that six runs, r22, r39, r58, r61, r70, and r82, behave unusually.

When runs are sorted by the final value in likelihood.dat, these six runs are ranked at the top because their last likelihood value is 0. However, their mean correlation with the final run r38 is only about 0.2299.

For these six runs, likelihood.dat ends with 0 / 1 followed by 0 / NaN. For example, r22/likelihood.dat ends with:

96  -1298817129.3705654144  -1.02547e-11
97   0.0000000000            1.00000e+00
98   0.0000000000           -nan

The same runs also show a late alpha jump in polarLDA.log. Alpha stays near 0.17-0.18 for many updates, then jumps to about 28712 and 58063. For example, r22 ends with:

alpha update 96: 0.17875
alpha update 97: 28712.55964
alpha update 98: 58063.63348

The final selected run r38 does not show this jump and finishes with alpha about 0.18406.

I also tried rerunning K=3 with a different seed offset, but a similar pattern still appeared in some runs. Would you have any suggestions on how to handle this case?Thanks in advance!

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions