IanMcDowell
commented
6 years ago I have done little work testing various concentration parameters. In practice, the number of clusters also strongly depends on a noise hyperprior. The defaults worked well for me for RNA-seq data but I widened the prior for microarray data, for example. I would recommend fixing sigma_n shape or rate and systematically varying the other hyperparameter.
On Wed, Mar 21, 2018 at 7:42 PM, jcharalel notifications@github.com wrote:
Hi Ian,
Do you have any recommendations on the alpha parameter? I see that the default is 1 but I was wondering if you have thoughts on what a reasonable range is? I have tried 0.25, 0.5, 1, 2, and considering going smaller as I feel like the cluster number would be better if further reduced (15-20 with the ones tried so far).
Thanks!
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://urldefense.proofpoint.com/v2/url?u=https-3A__github.com_PrincetonUniversity_DP-5FGP-5Fcluster_issues_8&d=DwMCaQ&c=imBPVzF25OnBgGmVOlcsiEgHoG1i6YHLR0Sj_gZ4adc&r=QGd0b_06v3hDpkoxZUB9c2iQ5RMTOts_DJ98uEZzA-o&m=tLzn0sNky83Qy9ZvM9QZsYYk-INjvcF2tIjTVHBFtdg&s=zGgahedg_85BP4jRtB24eTzhMjbDVOY_a-XLFlulAbY&e=, or mute the thread https://urldefense.proofpoint.com/v2/url?u=https-3A__github.com_notifications_unsubscribe-2Dauth_AFnc9CCgN7O5YqXWU-2DLrvoO57sSSuO8Gks5tguVjgaJpZM4S2OpX&d=DwMCaQ&c=imBPVzF25OnBgGmVOlcsiEgHoG1i6YHLR0Sj_gZ4adc&r=QGd0b_06v3hDpkoxZUB9c2iQ5RMTOts_DJ98uEZzA-o&m=tLzn0sNky83Qy9ZvM9QZsYYk-INjvcF2tIjTVHBFtdg&s=lYlphrlG4yhqYn2DDvomQ4eA3tp0r9zHg2-YgXRpxDA&e= .
-- Ian McDowell Bioinformatician II Duke University
jcharalel
Hi Ian,
Do you have any recommendations on the alpha parameter? I see that the default is 1 but I was wondering if you have thoughts on what a reasonable range is? I have tried 0.25, 0.5, 1, 2, and considering going smaller as I feel like the cluster number would be better if further reduced (15-20 with the ones tried so far).
Thanks!