Dear Tal, dear list,
I think the importance of interactive graphics has a lot do with how visual your
scientific discipline works. I'm spectroscopist, and I think we are very
visually oriented: if I think of a spectrum I mentally see a graph.
So for that kind of work, I need a lot of interaction (type: plot, change a bit,
plot again), e.g.
One example is the removal of spikes from Raman spectra (caused e.g. by cosmic
rays hitting the detector). It is fairly easy to compute a list of suspicious
signals. It is already much more complicated to find the actual beginning and
end of the spike. And it is really difficult not to have false positives by some
automatic procedure, because the spectra can look very different for different
samples. It would just take me far longer to find a computational description of
what is a spike than interactively accepting/rejecting the automatically marked
suspicions. Even though it feels like slave work ;-)
Roughly the same applies for the choice of pre-processing like baseline
correction. A number of different physical causes can produce different kinds of
baselines, and usually you don't know which process contributes to what extent.
In practice, experience suggests a method, I apply it and look whether the
result looks as expected. I'm not aware of any performance measure that would
indicate success here.
The next point where interaction is needed pops up as my data has e.g. spatial
and spectral dimensions. So do the models usually: e.g. in a PCA, the loadings
would usually capture the spectroscopic direction, whereas the scores belong to
the spatial domain. So I have "connected" graphs: the spatial distribution
(intensity map, score map, etc.), and the spectra (or loadings).
As soon as I have such connections I wish for interactive visualization:
I go back and forth between the plots: what is the spectrum that belongs to this
region of the map? Where on the sample are high intensities of this band? What
is the substance behind that: if it is x, the intensities at that other spectral
band should correlate. And then I want to compare this to the scatterplot (pairs
plot of the PCA score) or to a dendrogram of HCA...
Also, exploration is not just prerequisite for models, but it frequently is
already the very proper scientific work (particularly in basic science). The
more so, if you include exploring the models: Now, which of the bands are
actually used by my predictive models? Which samples do get their predictions
because of which spectral feature?
And, the "statistical outliers" may very well be just the interesting part of
the sample. And the outlier statistics cannot interprete the data in terms of
interesting ./. crap.
For presentation* of results, I personally think that most of the time a careful
selection of static graphs is much better than live interaction.
*The thing where you talk to an audience far awayf from your work computer. As
opposed to sitting down with your client/colleague and analysing the data together.
It could be argued that the interactive part is good for exploring (For
example) a different behavior of different groups/clusters in the data. But
when (in practice) I approached such situation, what I tended to do was to
run the relevant statistical procedures (and post-hoc tests)
As long as the relevant measure exists, sure.
Yet as a non-statistician, my work is focused on the physical/chemical
interpretation. Summary statistics are one set of tools for me, and interactive
visualisation is another set of tools (overlapping though).
I may want to subtract the influence of the overall unchanging sample matrix
(that would be the minimal intensity for each wavelength). But the minimum
spectrum is too noisy. So I use a quantile. Which one? Depends on the data. I'll
have a look at a series (say, the 2nd to 10th percentile) and decide trading off
noise and whether any new signals appear. I honestly think there's nothing
gained if I sit down and try to write a function scoring the similarity to the
minimum spectrum and the noise level: the more so as it just shifts the need for
a decision (How much noise outweighs what intensity of real signal being
subtracted?). It is a decision I need to take. With number or with eye. And
after all, my professional training was thought to enable me taking this
decision, and I'm paid (also) for being able to take this decision efficiently
(i.e. making a reasonably good choice within not too long time).
After all, it may also have to do with a complaint a colleague from a
computational data analysis group once had. He said the bad thing with us
spectroscopists is that our problems are either so easy that there's no fun in
solving them, or they are too hard to solve.
- and what I
found to be significant I would then plot with colors clearly dividing the
data to the relevant groups. From what I've seen, this is a safer approach
then "wondering around" the data (which could easily lead to data dredging
(were the scope of the multiple comparison needed for correction is not even
clear).
Sure, yet:
- Isn't that what validation was invented for (I mean with a proper, new,
[double] blind test set after you decided your parameters)?
- Summarizing a whole data set into a few numbers, without having looked at the
data itself may not be safe, either:
- The few comparisons shouldn't come at the cost of risking a bad modeling
modelling strategy and fitting parameters because the data was not properly
examined.
My 2 ct,
Claudia (who in practice warns far more frequently of multiple comparisons and
validation sets being compromised (not independent) than of too few data
exploration ;-) )
--
Claudia Beleites
Dipartimento dei Materiali e delle Risorse Naturali
Università degli Studi di Trieste
Via Alfonso Valerio 6/a
I-34127 Trieste
phone: +39 0 40 5 58-37 68
email: cbelei...@units.it
______________________________________________
R-help@r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
and provide commented, minimal, self-contained, reproducible code.
