Error Bars for plotting measures

[wingedRuslan]

Before adding error bars to the plot created by plot_network_measures function, I want to make sure that I create error bars correctly. (so that I will not need to change again code)

Here is an example of plotting error bar for just 1 measure:

So I take the mean value of random measures and then calculate the standard deviation of the random values.

While plotting I pass the mean value as height and standard deviation as an errorbar.

Is that the right way?

[KirstieJane]

Hi @wingedRuslan! This looks great.

Having just had a google around, I actually think showing the variance using the 95% confidence interval makes most sense (especially as for this plot the variance is so small!)

I think I'd code this up using the seaborn barplot function: https://seaborn.pydata.org/generated/seaborn.barplot.html. By default it will add the 95% CI error bars.

The work therefore is in building a data frame with the x and y values that you need. I don't have time to open up jupyter etc to figure out the code, but if you structure the data frame so that you can run hue="random" and then plot the real and random measures next to each other that should look great. (See for inspiration the plot from the documentation below).

>>> ax = sns.barplot(x="day", y="total_bill", hue="sex", data=tips)

The more I stare at this the more I wonder if the dataframe will have to be weirdly long to actually get the seaborn function working....so it might be better to go with something similar to what you've got above.

I can take a look tomorrow afternoon though if you have a bash 😄

[KirstieJane]

Here's a formula for the 95% confidence interval if you do stay with the code you've added at the top of this issue: 0.96SD to 1.05SD (https://www.graphpad.com/guides/prism/7/statistics/stat_confidence_interval_of_a_stand.htm?toc=0&printWindow)

[wingedRuslan]

Hi @KirstieJane

Thank you for tips!

This is the plot I have when there are only 5 random graphs:

By default seaborn.barplot adds the 95% CI error bars.

What do you think about the figure?

[KirstieJane]

This looks great!! I don’t like the colours (I like the grey for the random graphs) but the plot looks fantastic!

[KirstieJane]

Oh! And one other thing. I think the ratio can be a bit “squarer”. Change the figsize option to make the bigger number (the x dimension) a little smaller.

[wingedRuslan]

@KirstieJane I did not test for random_graphs = 1000 (or at least 100), as my system froze 😕
So tomorrow I want to run the code for a bigger number of random graphs and I expect to see confidence interval on other measures as well. Right now it is displayed only for assortativity and shortest_path measures.

Sure, I will change colors and make figure squarer :)

[wingedRuslan]

I've made the plotting function to leverage seaborn capabilities.
But building an input (dataframe) for seaborn.barplot takes a lot of time.

100 random graphs - 5 mins

@KirstieJane What is your opinion on that?

[KirstieJane]

@wingedRuslan can you split the timings into the "making random graphs", "network measures" and plotting steps? I suspect that 1 and 2 take a while, but the plotting should be very quick

[KirstieJane]

(plot looks great btw!)

[wingedRuslan]

@KirstieJane ohhhhh, yesterday was a bad day for me in terms of attention, so many small mistakes

you are right about time complexities for different operations!
Yesterday, I plotted the measures without having already calculated measures for 100 random graphs, that's why it took 5 mins, as plotting function was calculating these values.

If we pass GraphBundle with already calculated measures, it is less than 1 sec for 100 random graphs.

yeah, finally I addressed all the comments and ready to submit my commits to PR #121