5. Visualization: every plot, and how to read it

Horizontal phylogram (default)

Root on the left, depth rightward; every leaf is labelled with its full arrow-form pathway and the predicted next state. This is the style for a paper when you need to cite specific pathways inline.

plot(pruned, style = "horizontal")

point_size_range and edge_size_range exaggerate or compress the size dynamic range – useful for slides where the count contrast must read from the back of the room. The encodings are unchanged; only the scales differ.

plot(pruned, style = "horizontal",
     point_size_range = c(3, 12), edge_size_range = c(0.4, 3.5))

Radial dendrogram

The same tree wrapped into a circle: the eye goes to the thick central branches (the corpus highways) versus the thin outer twigs (contexts pruning kept on evidence, not volume).

plot(pruned, style = "dendrogram")

Icicle / sunburst

A space-filling partition: arc angular width is proportional to count, so a dominant state visually swallows the ring – an honest depiction of class imbalance.

plot(pruned, style = "icicle")

A fourth style, style = "interactive", renders the same tree as a draggable, zoomable visNetwork widget (collapse the dominant spine and the rare informative branches become legible). It produces an HTML widget rather than a static figure, so it is best run in an interactive session rather than shown inline here.

Next-state heatmap

Each row is a context, each column a next state, each cell P(next | context), modal cell bold; a > prefix marks a context whose modal next state flips versus its shorter parent. Sorting the same data two ways is the single best “common vs informative” figure:

plot_pathways(pruned, top = 12, sort_by = "count")        # the highways

plot_pathways(pruned, top = 12, sort_by = "divergence")   # the informative ones

Sorted by count the bright cells stack on the most frequent next state; sorted by divergence they move off it. That lateral shift is the thesis in one comparison.

Divergence lollipop

Per-context KL from the shorter parent, ranked, with orange points marking modal-flip contexts – the histories that genuinely change the prediction. min_count removes small-sample mirages.

plot_divergence(pruned, top = 12, min_count = 5)

Per-context distributions

The full next-state distribution for each context as small multiples – peaked panels are near-settled continuations, flat panels are the decision points where history does not resolve the next state.

plot_distributions(pruned, top = 6)

How much memory does one pathway need?

plot_pruning() walks a pathway’s suffix chain – the full context, then the same context with its oldest move dropped, down to the root – and marks which contexts the pruning test keeps (solid) versus drops (faded). It answers, for that one pathway, how far back history actually has to reach.

plot_pruning(tree, "Active -> Active -> Average")

Predictive quality

plot_predictive() scores sequences against the fitted tree three ways. For this tour we score the bundled trajectories themselves; in a real evaluation pass genuinely held-out sequences (the Advanced analysis vignette shows the cross-validated route).

type = "logloss" – per-position surprise in bits against position; below the uniform ceiling is structure the model exploited:

plot_predictive(pruned, trajectories, type = "logloss")

type = "ecdf" – the distribution of the probability assigned to the state that actually occurred; steep steps reveal calibration plateaus (e.g. a mass of three-way-open branch points):

plot_predictive(pruned, trajectories, type = "ecdf")

A third type, type = "position", traces each individual sequence’s confidence move-by-move (one grey line per sequence). It is a per-sequence view that only reads cleanly for a handful of sequences, so it is omitted here; reach for it when you want to inspect a few specific trajectories rather than the corpus as a whole.

Bootstrap forest plot

Each pathway’s 95% bootstrap interval on G-squared against the chi-square critical value (dashed line); colour encodes the trust quadrant. A bar entirely to the right is reproducibly informative.

boot <- bootstrap_pathways(pruned, iter = 100L, seed = 1L)
plot(boot)
#> `height` was translated to `width`.

Per-pathway resample distributions

plot_pathway_resamples(boot, stat = "divergence", top = 6)

Cohort comparison: permutation null

We name an external group column (Achiever) on the bundled group_regulation_long log; context_tree(group = ) fits one tree per cohort, and compare_trees() consumes the group directly.

data(group_regulation_long)
grp_reg <- context_tree(group_regulation_long,
                       actor = "Actor", time = "Time", action = "Action",
                       group = "Achiever", max_depth = 2L, min_count = 10L)
cmp <- compare_trees(prune_tree(grp_reg), iter = 199L, seed = 1L)
plot(cmp)

The observed distance (orange line) sits in the right tail of the label-shuffled null (grey) – the visual form of the permutation p-value.

Tuning surface

tg <- tune_tree(trajectories, max_depth = 1L:4L, folds = 5L, seed = 1L)
plot(tg)

A flat-then-rising perplexity curve is the picture of a short-memory process; the orange star marks the cross-validated winner.

Group difference map

plot_difference() draws the per-context residual map for the same group =-fitted tree – where two cohorts resolve the same history toward different next states. depth = 1L keeps the map to the single-state contexts so the rows stay legible (a deep tree has too many contexts to label).

plot_difference(grp_reg, depth = 1L)