All five vignettes now render their figures at dpi = 110. Three were at 150, which pushed the source tarball past CRAN’s 5 MB guidance; the package is materially smaller as a result (installed size 7.3 Mb to 5.8 Mb, of which doc falls from 5.9 Mb to 4.5 Mb) with no change to any reported number.
The tutorial test files pin the bootstrap edge identities, the high -> moderate and low -> low edge weights, the low -> low p-value and the low state share, so a change that reshuffled which edges survive resampling can no longer pass while contradicting the vignette prose.
tsn 1.4.0
Breaking changes
The bundled example data sets have been replaced. steps and motivation (which carried no usable provenance) are gone; in their place the package ships two data sets with full provenance owned by the package authors:
srl — a balanced daily panel of 36 students reporting nine self-regulated-learning indicators for 156 occasions on 0-100 scales, the companion data of Advanced Learning Analytics Methods (CC BY-NC-SA 4.0; see the COPYRIGHTS file). Identical to the copy in the authors’ ‘idiographic’ package.
esm_srl — the authors’ own fully anonymized experience-sampling study: 41 students, momentary self-regulation/motivation/anxiety on 0-100 scales, fictional identifiers and constant-shifted dates. The tsn copy adds a derived day_type (weekday/weekend) factor so grouped models need no caller-side preparation. Code that referenced steps or motivation must switch data sets.
The internal R/sysdata.rda copy of the old data sets (an unused leftover that silently re-shipped them inside the namespace) is removed.
Documentation
Every vignette, example, and the README were rebuilt on the new data, with all reported numbers recomputed:
vignette("nestimate-workflow") runs on the srl effort panel (fixed breaks at 40/70), one student’s series for segmentation, a first-versus-second-half course comparison, and the weekday/weekend grouped model.
vignette("group-models") compares weekday and weekend effort dynamics end to end, including the welding and threshold failure modes of hand-subsetting.
Grouped transition models: ts_tna(), ts_ftna(), ts_cna(), and ts_atna() gain a group argument naming a column of data. The result is one network per group — a ts_tna_group (also a Nestimate netobject_group) that Nestimate’s compatible grouped verbs (net_prune(), state_distribution(), net_centrality(), compare_model(), permutation()) accept directly. States are discretized from the pooled series before the split, so every group shares one alphabet and node set. A sequence never spans a group boundary: contiguous runs of the group column become their own sequences, so no transition is ever counted between observations that were not adjacent in time. as.data.frame() on the collection returns tidy edges, series, or groups tables.
segment and overlap on the ts_*na() family cut each series into shorter sequences so sequence-based inference (bootstrap, permutation, stability) has units to resample. Partitioned blocks are the conservative default; overlap = TRUE slides the block one observation at a time and at segment = 2 reproduces the unsegmented estimate exactly.
New as.matrix.ts_tna() and as.data.frame.ts_tna() methods: the weight matrix and a tidy one-row-per-state-pair edge table (plus what = "series" for the per-observation source table).
New vignette("nestimate-workflow"): raw series to transition network to the Nestimate inferential stack, including segmentation and grouped comparison.
New vignette("group-models"): a full tutorial on grouped transition models — why hand-subsetting both fabricates transitions and shifts the state thresholds, how the pooled alphabet and boundary rule fix it, and the descriptive and inferential comparison of groups end to end.
New vignette("nestimate-compatibility"): the verb-by-verb map of which Nestimate analyses accept a tsn model directly — reading, composition, dynamics, entropy, pruning, reliability, and model comparison — plus the verbs that do not apply and why, ending in a compatibility table.
New plot.ts_tna_group() method: draws one selected network of a grouped model (plot(x, group = "Home")) with the full plot.ts_tna() surface, so a group can be rendered without reaching into the collection.
Breaking changes
Arguments that the selected configuration cannot consume are now rejected with an error instead of being silently ignored — for example chain or normalize on a visibility network, limit, decay, penetrable, or aggregation on a distance network, step outside unit = "window", p without distance = "minkowski", breaks without discretization = "threshold", n_states with the ordinal discretizer, seed with a deterministic discretizer, and any discretization option when data is already a tsn_states. Shortcut methods ("nvg", "hvg", discretizer names) likewise reject granular arguments that contradict what the shortcut implies. Previously these calls returned a plausible but unintended model.
summary() density now uses a fixed possible-edge universe per network family: state networks always count self-loops among the possible edges, and series/window/time networks never do. Previously the denominator depended on whether a loop happened to be observed, so densities were not comparable across models. State networks without observed loops report a lower (correct) density than before.
Caller-supplied factor state values now fix the node order by their level order, unused factor levels are retained as zero-degree nodes, and the stored source state column stays a factor. Discretized states order nodes numerically ("1", "2", …) instead of by first appearance. Plain character states keep first-appearance order.
The "okabe" plot palette now rejects requests for more than its nine colour-blind-safe colours instead of returning NA colours.
Bug fixes
series selection on long data now follows the requested series order, exactly as wide, matrix, and list selection do. Previously the same selection could reverse chained or directed edges depending on the input container.
Point-level distance networks (method = "distance", unit = "time") now use the same collision-safe node labels as the visibility family, so valid (id, time) pairs whose display strings collide are no longer rejected.
ts_*na(series = ..., group = ...) no longer warns that groups excluded by the caller’s own series selection were “lost to discretization”.
Performance
Natural visibility construction now runs in quadratic instead of cubic time (and no longer materializes all candidate pairs at once): a 1,000-observation series builds in well under a second. The same scan accelerates horizontal visibility with penetrable or limit set.
tsn 1.2.0
series_networks() now returns a tsn_series_networks collection instead of a bare list. print() and summary() give a tidy one-row-per-series table (series, type, observations, states, edges), as.data.frame() returns the same table as a data frame, and plot() draws one model directly. series selects the model by name; it may be omitted only when the collection holds exactly one model.
plot() on a network result now applies restrained presentation defaults (spring layout, nodes scaled by degree, one node fill, muted edges, no edge labels) so a bare plot(network) is readable. Every default is a named cograph::splot() argument and any value supplied through ... overrides it, so the full cograph surface is unchanged.
The overlay default for state and source-series views is now "horizontal" (value bands) rather than "vertical" (time runs), affecting plot.tsn_states(), plot(type = "series"), and plot.ts_tna(). Pass overlay = "vertical" to restore the previous shading.
New vignette("plotting-time-series-networks"): the plotting surface end to end — network and source-series views, state overlays, ribbon, heatmap, and stack types, and transition-model plots.
vignette("pleasure-all-functions") was rewritten around the packaged motivation pleasure series, selecting it through each verb’s series argument rather than extracting the column.
tsn 1.1.0
New vignette("pleasure-all-functions"): a complete tutorial using the packaged motivation$pleasure measurements to demonstrate all nine exported functions, the standard result interface, and cograph-only network plots.
New Nestimate bridge: ts_tna(), ts_ftna(), ts_cna(), and ts_atna() discretize one or many time series and build the matching Nestimate transition-network model (netobject) that keeps the source data ($ts_source) — plottable and compatible with Nestimate analyses whose sampling requirements the data satisfy. Node order follows the state level order (low/mid/high), not alphabetical.
New series_networks() verb: splits a pooled bridge model into one full model per source series, all sharing the same node set and order for direct comparison and retaining the original Nestimate builder settings.
plot() on a bridge result draws each series beside a transition network on the same row — by default the series’ own network (network = "per_series"); network = "summary" draws the pooled model spanning all rows instead. Networks render exclusively through cograph::splot(), with state-matched node fills and nodes sized by in-strength (self-loops excluded); ... passes straight to splot.
New time-series plot types, all base graphics. plot() on a discretize() result accepts type = "ribbon" (clean series with a state-classification strip underneath), "heatmap" (one row per series, one tile per observation — compares state timing across many series), and "stack" (a companion series such as the original values stacked above the discretized series, both panels shaded by the discretized states, with optional state-coloured line segments via color_line), alongside the previous shaded view (now type = "overlay", still the default). The overlay view combines colour shading in one direction with dashed guide lines in the other via lines = ("horizontal" at the state boundaries, "vertical" at persistent state transitions filtered by min_run). plot() on a trend() result accepts "ribbon", "heatmap", and "panels" (each series stacked above the rolling metric it was classified on), alongside the previous default (type = "points").
Five new distance measures for tsn(method = "distance"), all base-R with no new dependencies: "ccf" (one minus the maximum absolute cross-correlation across lags), "nmi" (one minus normalized mutual information on shared quantile bins), "voi" (variation of information), "event_sync" (one minus the Quiroga event-synchronization index over event times), and "van_rossum" (exact closed-form van Rossum spike-train distance). New method-specific arguments: bins (nmi/voi), lag (ccf), and tolerance (an optional cap on adaptive event-synchronization windows, or the van Rossum time constant). NMI and VoI are verified equivalent to aricode; the van Rossum exact form is verified against the discretized-grid reference.
Visibility geometry, edge distances, temporal limits, and decay now use the supplied numeric or date-time axis; nonnumeric labels use observation steps. Collision-safe node labels preserve distinct (id, time) pairs.
Temporal discretizers now compute ordinal patterns, adaptive-magnitude features, and DTW windows within each series before fitting a shared state vocabulary, preventing patterns from crossing ID boundaries.
Ordinal transition builders exclude trailing display-fill labels, and NMI/VoI now use marginal quantile bins so monotone rescaling preserves the expected information distance.
Distance chains preserve caller series order and never link the last window of one series to the first window of another.
Event synchronization now implements the adaptive Quian Quiroga–Kreuz– Grassberger coincidence window and validates strictly increasing event times.
trend() now supports missing observations, respects numeric/date elapsed time, handles character time without coercion warnings, avoids infinite growth factors, and validates its classification thresholds.
Plot guides now map transformed and magnitude-space breaks back to the raw signed axis; cograph transition-network node sizing excludes self-loops, and show_weights = FALSE is honoured.
New similarity argument selecting the distance-to-weight kernel: "inverse" (the previous fixed rule), "normalized_inverse", "negative_exp", and "gaussian", the latter two scaled by bandwidth (defaulting to the median positive distance).
normalize now also accepts mode names: TRUE/"max" (previous behaviour), "minmax", and "quantile" (5th-95th percentile winsorized scaling).
Full upstream coverage: ported trend() (rolling OLS, Theil-Sen, Spearman, Kendall, and growth-factor trend classification with an epsilon flat band, turbulence override, and center/right/left alignment) and a public discretize() verb with custom labels and log/zscore transforms.
Added the vg() verb for visibility graphs: vg(x) and vg(x, "horizontal") are discoverable, named-argument wrappers around tsn(method = "visibility").
Results are now dual-class c("tsn", "netobject", "cograph_network") objects. With cograph installed, cograph::splot(x) renders any tsn network directly; as.data.frame(x) still returns the tidy dyad table and as.data.frame(x, what = "series") the source series.
Network plots now delegate exclusively to cograph::splot(); tsn no longer maintains a second generic graph renderer. plot(x, "series") retains the package-specific source-series view.
Method-string shortcuts: tsn(x, "hvg"), tsn(x, "nvg"), tsn(x, "distance"), and every discretizer name (tsn(x, "ordinal"), tsn(x, "quantile"), …) as a direct method on a bare numeric vector.
Fifteen discretizers reachable through both tsn(unit = "state") and discretize(), with identical state assignments.
New network options: chain (consecutive-only transition chains), directed distance networks, and normalize.
tsn 1.0.0
Rebuilt the package around the single tsn() entry point.
Added whole-series and sliding-window distance networks.
Added base-R natural, horizontal, and penetrable visibility graphs.
Added ten base-R distance measures, including unequal-length DTW.
Added five consistent distance-to-network connection rules.
Added seven internal state discretizers, including a base-R Gaussian mixture.
Standardized every result behind the tsn class with tidy dyad-table access.
Removed all external runtime dependencies and overlapping network verbs.
Corrected legacy selection, similarity direction, state ordering, DTW, visibility duplication, and interleaved-ID defects.
