Examples: Special Features

source("lkt-vignette-setup.R")
## Executable examples are disabled for routine package builds. Set LKT_RUN_EXPENSIVE_VIGNETTES=true to run this vignette.
knitr::opts_chunk$set(eval = run_expensive_vignette)

These full-data feature experiments are intentionally not evaluated during routine CRAN checks. Set LKT_RUN_EXPENSIVE_VIGNETTES=true to execute them when rendering.

Load data

val <- prepare_largeraw_sample()

brpropdec

model_brpropdec <- LKT(
  data = val,
  interc = TRUE,
  dualfit = TRUE,
  components = c("KC..Default.", "Anon.Student.Id", "KC..Default.", "KC..Default."),
  features = c("baseratepropdec", "logitdec", "logitdec", "recency"),
  fixedpars = c(0.988209, 0.9690458, 0.9004974, 0.2603806)
)

print(model_brpropdec$coefs)
check_lkt_fit(model_brpropdec, expected_r2 = 0.240340, expected_loglike = -28854.5108157355)
check_true("brpropdec latency model returned", !is.null(model_brpropdec$latencymodel))
check_has_coefficients(model_brpropdec, c("baseratepropdecKC..Default."))

brpropdec_spec <- model_brpropdec$model_specification[[1]]
check_true("duplicate logitdec specs are retained", all(c(
  "Anon.Student.Id",
  "KC..Default."
) %in% brpropdec_spec$component[brpropdec_spec$feature == "logitdec"]))
check_true("duplicate logitdec parameters stay component-specific", all(c(
  0.9690458,
  0.9004974
) %in% brpropdec_spec$para[brpropdec_spec$feature == "logitdec"]))

Simple adaptive model for practice optimization

model_adaptive <- LKT(
  data = val,
  interc = FALSE,
  dualfit = FALSE,
  factrv = 1e11,
  components = c(
    "Anon.Student.Id",
    "KC..Default.",
    "KC..Default.",
    "KC..Default."
  ),
  features = c("logitdec", "logsuc", "recency", "intercept"),
  fixedpars = c(0.98, 0.24)
)
check_lkt_fit(model_adaptive, expected_r2 = 0.317237, expected_loglike = -25933.7244347462)
check_has_coefficients(model_adaptive, c("recencyKC..Default.", "logsucKC..Default."))

adaptive_spec <- model_adaptive$model_specification[[1]]
check_true("adaptive model specification is tabular", data.table::is.data.table(adaptive_spec))
check_true("adaptive model specification has coefficient names", all(!is.na(adaptive_spec$coefficient_name)))
check_true("adaptive model specification matches coefficient rows", nrow(adaptive_spec) == nrow(model_adaptive$coefs))
check_true("adaptive model specification exposes PolicyFactory columns", all(c(
  "coefficient_name",
  "feature",
  "component",
  "component_level",
  "coefficient",
  "para",
  "parb",
  "parc",
  "pard",
  "pare"
) %in% names(adaptive_spec)))
adaptive_spec_filtered <- adaptive_spec[
  !(feature == "intercept" & component == "KC..Default.")
]
check_true("adaptive model specification supports KC intercept filtering", nrow(adaptive_spec_filtered) < nrow(adaptive_spec))
check_true("adaptive model specification stores logitdec parameter", any(
  adaptive_spec$feature == "logitdec" &
    adaptive_spec$component == "Anon.Student.Id" &
    adaptive_spec$para == 0.98
))
check_true("adaptive model specification stores recency parameter", any(
  adaptive_spec$feature == "recency" &
    adaptive_spec$component == "KC..Default." &
    adaptive_spec$para == 0.24
))

KC intercepts across time

val_time <- val[order(val$CF..Time.), ]
model_kc_time <- LKT(
  data = val_time,
  interc = TRUE,
  dualfit = FALSE,
  factrv = 1e11,
  components = c(
    "Anon.Student.Id",
    "KC..Default.",
    "KC..Default.",
    "KC..Default."
  ),
  features = c("logitdec", "logsuc", "recency", "logitdecevol"),
  fixedpars = c(0.98, 0.24, .99)
)
check_lkt_fit(model_kc_time, expected_r2 = 0.305281, expected_loglike = -26387.8587874768)
check_has_coefficients(model_kc_time, c("logitdecevolKC..Default."))

Astonishing model

model_astonishing <- LKT(
  data = val,
  interc = TRUE,
  dualfit = TRUE,
  factrv = 1e7,
  components = c(
    "Anon.Student.Id",
    "KC..Default.",
    "KC..Default.",
    "KC..Default."
  ),
  features = c("intercept", "intercept", "lineafm$", "lineafm"),
  interacts = c(NA, NA, NA, "Anon.Student.Id")
)
check_lkt_fit(model_astonishing, expected_r2 = 0.309291, expected_loglike = -26235.522445)
check_true("astonishing latency model returned", !is.null(model_astonishing$latencymodel))
check_has_coefficient_matching(model_astonishing, "^lineafmKC[.][.]Default[.]:Anon[.]Student[.]Id")

Build LKT with special feature

model_search_special <- buildLKTModel(
  data = val,
  interc = TRUE,
  specialcomponents = "CF..End.Latency.",
  specialfeatures = "numer",
  allcomponents = c("Anon.Student.Id", "KC..Default."),
  currentcomponents = c(),
  forv = 100,
  bacv = 80,
  allfeatures = c("lineafm", "logafm", "logsuc", "logfail", "linesuc", "linefail"),
  currentfeatures = c(),
  currentfixedpars = c(),
  forward = TRUE,
  backward = TRUE,
  maxitv = 1,
  verbose = FALSE
)
check_true("special feature search returns table and model", length(model_search_special) == 2)
check_true("special feature search has selected model", !is.null(model_search_special[[2]]$coefs))