Changelog#

v0.1.0 (2026-11-12)#

Feat#

  • data: medoids, coincidence factor and cluster profile panels in the book style (#35)

  • cluster: row standardisation, clustering by correlation and one-to-one naming of clusters by reference classes (#34)

  • features: range scaler, exposure of units to events and group mean profiles; NMI in the ablation measure

  • features: event map with circular time of day and year

  • cluster: variance share with its permutation null and bias correction

  • plot: cleaner cluster scatter with hollow-edged markers, trimmed outlines and repelled numbers

  • plot: cluster scatter with markers, numbers, outlines and a highlighted cluster

  • cluster: stable weighted Borda rank (#27)

  • cluster: normalisation forms, Hopkins tendency and a minimum cluster size rule (#26)

  • plot: colour accessibility measures (#25)

  • a single-day clock error and the representations of a day (#24)

  • alternatives for comparing loss models, attribution, detectors and monitors (#23)

  • anomaly: threshold rules, a reconstruction detector, point adjustment and day features (#22)

  • anomaly: collective export of a group of meters (#21)

  • benchmark: Wilcoxon signed-rank test of paired results (#20)

  • evaluate: behaviour faults, adoption, monitoring and faulty case bases (#19)

  • anomaly: energy balance of a transformer with technical-loss models (#18)

  • CUSUM paths with a calibrated limit, physical inverter faults, adoption and meter validation (#17)

  • features: circular timing option for event features (#14)

  • anomaly: control-limit detector, conformal p-values and false-discovery-rate selection (#13)

  • recommend: ranking metrics, learning to rank, equity and uplift (#8)

  • plot: replaceable lets-plot style through theme.use (#6)

  • safina v0.1, anomaly detection, clustering and retrieval as scikit-learn estimators (#1)

Fix#

  • release: reclaim ownership of the workspace after the Docker bump so the changelog can be written (#41)

  • publish the distribution as pundamilia, keeping milia as the import name (#40)

  • cluster: tolerate float ties in the permutation p-value and test the variance share against the ANOVA identity

  • use day_of_week and day_of_year, not the deprecated pandas aliases (#16)

  • plot: sentence-case every displayed label (#7)

  • plot: readable faceted bars, square scatters, integer heatmap cells; seed fANOVA (#4)

Refactor#

  • rename the package from safina to milia (#39)

  • cluster: name the variance share explained, as the naming rule requires

Perf#

  • features: vectorise event features over a set of days (#15)