scaling of factors in plot_type = "scatter"

[stephens999]

it seems in the vignette (flashier_single_cell.html) scatter plots the factors are normalized
to have max value 1, but I don't think we necessarily want or need that scaling?

(The heatmap is also scaled that way, but arguably this could be helpful in the heatmap because we are showing all the factors at once.)

[pcarbo]

@stephens999 Are you suggesting to plot L_pm or F_pm? Currently I think @willwerscheid uses ldf() to normalize the factors.

[stephens999]

I am suggesting to plot DF' instead of F' (with L normalized with infinity norm).
The vignette does print out the values of DF' for the top genes, but it doesn't include D in the plot.

[pcarbo]

Yes, that makes sense.

This reminds me that I wanted to make a few related improvements to the plotting functions in flashier, mainly to clarify what is being plotted.

[pcarbo]

@stephens999 @willwerscheid I've been thinking about how to make the plots more consistent and uniform in interpretation:

plot_type = c("scree", "bar", "heatmap", "histogram", "scatter", "structure")

(Aside from the screen plot which is different.)

What do you think about the following conventions for the plots:

• If a single k is being plotted, show with(ldf(fl,type = "i"),F %*% diag(D)) or `with(ldf(fl,type = "i"),L %*% diag(D)).

• If more than one k is being plotted, show ldf(fl,type = "i")$F or ldf(fl, type = "i")$L so that the values are more comparable across k.

• Potentially we could have a TRUE/FALSE option (e.g., "normalize") to control this behaviour.

[stephens999]

I think I prefer to stick to a default of putting F with D, and L alone: so with(ldf(fl,type = "i"),F %*% diag(D)) and with(ldf(fl,type = "i"),L) Matthew

…

On Tue, Jul 30, 2024 at 12:09 PM Peter Carbonetto ***@***.***> wrote: @stephens999 <https://github.com/stephens999> @willwerscheid <https://github.com/willwerscheid> I've been thinking about how to make the plots more consistent and uniform in interpretation: plot_type = c("scree", "bar", "heatmap", "histogram", "scatter", "structure") (Aside from the screen plot which is different.) What do you think about the following conventions for the plots: - If a single k is being plotted, show with(ldf(fl,type = "i"),F %*% diag(D)) or `with(ldf(fl,type = "i"),L %*% diag(D)). - If more than one k is being plotted, show ldf(fl,type = "i")$F or ``ldf(fl, type = "i")$L` so that the values are more comparable across k. - Potentially we could have a TRUE/FALSE option (e.g., "normalize") to control this behaviour. — Reply to this email directly, view it on GitHub <#131 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AANXRRJNWBB3UAIVJDPF4QTZO7QHTAVCNFSM6AAAAABLM5OCWGVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDENJZGAZDONJQHA> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[pcarbo]

@stephens999 Okay, but it sounds like this would only be the default when plotting a single k? As you said previously,

The heatmap is also scaled that way [that is, normalized so that the maximum value is 1], but arguably this could be helpful in the heatmap because we are showing all the factors at once.

In other words, the default will be with(ldf(fl,type = "i"),F %*% diag(D)) only when plotting a single column of F.

[stephens999]

i think that for simplicity the default for the heatmap could be the same (DF'), but that user might in some cases (when D differs a lot across factors) want to override that default. (There are arguments both ways - even when D differs a lot, one might want to see that information in the heatmap...) Matthew

…

On Tue, Jul 30, 2024 at 10:28 PM Peter Carbonetto ***@***.***> wrote: @stephens999 <https://github.com/stephens999> Okay, but it sounds like this would only be the default when plotting a single k? As you said previously, The heatmap is also scaled that way [that is, normalized so that the maximum value is 1], but arguably this could be helpful in the heatmap because we are showing all the factors at once. In other words, the default will be with(ldf(fl,type = "i"),F %*% diag(D)) only when plotting a single column of F. — Reply to this email directly, view it on GitHub <#131 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AANXRRIKWLASOOTIDT2RBEDZPBKXVAVCNFSM6AAAAABLM5OCWGVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDENJZGU3DQNRYGE> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[pcarbo]

@willwerscheid Matthew is suggesting the following default for all plots:

• When plotting L, ldf(fl, type = "i")$L is shown.

• When plotting F, with(ldf(fl,type = "i"),F %*% diag(D)) is shown.

Are you okay with this as the default setting for all plots? Are there particular plots where you think this default might be problematic?

If you are okay with this, I will update the plotting interface and propose the changes in a pull request.

[pcarbo]

I created a branch "improvements_to_plots" to address this issue and tackle other related improvements.