SBR score plot: marker area carries the SPE, batch 4 labelled above - #91
Conversation
…led above The score plot showed only how extreme a batch is along the components. Its marker area is now the batch's SPE, so one figure answers both questions asked of a batch: batch 37, the furthest out along the first component, has the smallest residual of the 53 batches, and batch 34 the twelfth smallest, which is the point the influence plot beside it then makes with axes. Area, not diameter, carries the value, and two grey circles in the legend give the scale, without which an area cannot be read off a plot. A highlighted batch keeps its own area and is marked by a heavier edge instead of a larger one, because two meanings on one channel cannot both be read. `score_plot` also takes `label_north`, for a label centred above its marker rather than to its right; a label placed inside a cloud gets a white outline so it reads over whatever it crosses. Batch 4 uses both. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LXthGpHLQFfGubBiKFtGAE
|
CI on this head: "The colour case study, with pyoptex" passes, "Against process-improve main (advisory)" is still running, and "Against the PyPI release" fails, on the same blocker as #84 rather than on anything in this change. That job installs The three failures are the three batch scripts, and they were failing this way before this PR. Nothing in this repository fixes it: the job goes green when The change itself was checked by regenerating the whole SBR figure set against the merged library: one PNG differs, Generated by Claude Code |
… read clearly The batches span a factor of two in SPE, which as an area is a factor of two and hard to judge across a scatter. The area now carries the square of the SPE, a factor of four, and the sum of squared residuals is in any case the quantity that adds up over the cells of a batch. The legend key stays in the reader's units: `score_plot` takes `size_of_reference`, which maps a reference value onto whatever scale the area encodes, so the circle labelled "SPE 20" is the size a batch with an SPE of 20 is drawn. Three circles now, at 20, 30 and 40, spanning the batches. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LXthGpHLQFfGubBiKFtGAE
…er the grid The scatter about the y = x line is what a reader judges on a parity plot, so the number for it belongs on the same axes. Each panel's legend now lists RMSEE, the scatter of the fitted values, and RMSEP, the same scatter when every batch is left out of the fit in turn, each in the attribute's units and in standard deviations of it. The legend sits over the grid and close to the cloud, so it is drawn on white at 85% opacity with no frame: readable, without reading as a panel of its own. One leave-one-batch-out sweep now serves both figures that need it, its end-of-batch row here and the whole frame in the prediction-error curve later, rather than the script paying for it twice. The classical centre and spread in the departure figure are renamed, since they shadowed the quality standard deviations once that sweep moved above them. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LXthGpHLQFfGubBiKFtGAE
The RMSEP in the legend is a number; the band is the same thing as a distance, so a reader can see at a glance which batches the model misses by more than it should. Two RMSEP either side of y = x, drawn below the markers and above the grid, with the band rather than the outermost batch setting the view. `parity_plot` takes `band_from`, naming which of the errors it was given to shade, and `band_multiple`. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LXthGpHLQFfGubBiKFtGAE
A legend drawn straight onto the panel has the grid, and sometimes the data, running through its text. The shared style now gives every legend in these figures a white ground at 85% opacity and no edge, so it reads wherever it sits without becoming a panel of its own, and the parity plot drops the explicit keywords it needed before. Twenty-three figures across the three case studies change, in their legends only. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LXthGpHLQFfGubBiKFtGAE
The three docstrings asked for 1.79 or 1.81, neither of which was ever published; PyPI stands at 1.77.0. The loaders, the on-line methods and the multiblock pieces these scripts import all ship in 1.82.0. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LXthGpHLQFfGubBiKFtGAE
|
Correcting two numbers in the comment above, now that the release version is settled. "Against the PyPI release" fails on this head for the same reason and with the same tally ( The last commit here updates the three script docstrings accordingly: they asked for "version 1.79 or later" and "1.81 or later", neither of which a reader could install. The job is deterministic, so there is nothing to re-run: it goes green when 1.82.0 reaches PyPI, which also clears the two red checks on kgdunn/pid-book#273. Generated by Claude Code |
2 participants
Follow-up to #84, on review feedback for the SBR score plot (
batch/batch-case-sbr-scores.png).What changed
How it is drawn
BUBBLEis the area given to the median batch), so a marker of twice the area stands for twice the SPE.score_plotgainssizes,size_nameandsize_referencefor the encoding andlabel_northfor the label position;group_scattergainsareas. All are opt-in, so the DuPont and FMC figures are unchanged: regenerating the whole SBR set produces one changed PNG,batch-case-sbr-scores.png.The book page's Plotly code is updated to match in kgdunn/pid-book#273, where the caption now states what the marker area means, and the prose quotes the two SPE ranks (echoed in the code, so the book's checker compares them).
🤖 Generated with Claude Code
https://claude.ai/code/session_01LXthGpHLQFfGubBiKFtGAE
Generated by Claude Code