Getting started with segen
In segen: Sequence Generalization Through Similarity Network

knitr::opts_chunk$set(collapse = TRUE, comment = "#>", fig.width = 7, fig.height = 4)
library(segen)

segen forecasts numeric or categorical sequences by comparing recent windows with historical windows and aggregating similar sequences.

Numeric example

The package includes a small example dataset. The following deliberately uses one validation window and one sampled model so that the vignette runs quickly.

set.seed(123)
fit <- segen(
  time_features[, "IBM.Close", drop = FALSE],
  seq_len = 12,
  similarity = 0.7,
  n_windows = 2,
  n_samp = 1,
  seed = 123
)

head(fit$history)
fit$best_model$predictions$IBM.Close

The result contains the model search history, the selected model's predictions and testing errors, plots, and elapsed-time information.

fit$best_model$plots$IBM.Close

Reproducibility

Set seed whenever results need to be reproduced. This controls model sampling and uncertainty draws. segen runs sequentially by default.

Distance methods and parallel execution

All six distance methods, including DTW with the symmetric2 recurrence, are implemented using standard R libraries. Parallel execution uses PSOCK workers from the standard parallel package, with two workers by default.

fit_parallel <- segen(time_features[, "IBM.Close", drop = FALSE],
                     seq_len = 12, n_samp = 4,
                     use_parallel = TRUE, parallel_workers = 2)

Changes in 2.0.1

There are no contributed runtime dependencies. Plot objects now have class segen_plot; display them with plot() or print(), rather than adding ggplot2 layers. Numeric gaps use linear interpolation with constant endpoint extension; categorical gaps use the most frequent observed level. Smoothing uses degree-one loess with span 0.75 rather than automatic span selection. These preprocessing changes can change forecasts. Binary entropy now measures the empirical distribution of zero/one outcomes in natural-log units. Percentage metrics use percentages and denominators are bounded below by 1e-8. Intervals use pooled rolling residuals and do not guarantee nominal coverage under arbitrary temporal dependence.

Input requirements

Input must be a data frame whose columns are all numeric or all categorical. Dates, when supplied, must be a Date vector with one value per row. Missing numeric values are imputed before forecasting.