bayeswatch {bayesWatch}

Fit a bayesWatch object.

Description

Main method of package. MCMC sampling for change-point probabilities with fault detection according to the model by Murph et al. 2023. Creates a bayesWatch object for analysis of change-points.

Usage

bayeswatch(
  data_woTimeValues,
  time_of_observations,
  time_points,
  variable_names = 1:ncol(data_woTimeValues),
  not.cont = NULL,
  iterations = 100,
  burnin = floor(iterations/2),
  lower_bounds = NULL,
  upper_bounds = NULL,
  ordinal_indicators = NULL,
  list_of_ordinal_levels = NULL,
  categorical_indicators = NULL,
  previous_states = NULL,
  previous_model_fits = NULL,
  linger_parameter = 500,
  move_parameter = 100,
  g.prior = 0.2,
  set_G = NULL,
  wishart_df_initial = 1500,
  lambda = 1500,
  g_sampling_distribution = NULL,
  n.cores = 1,
  scaleMatrix = NULL,
  allow_for_mixture_models = FALSE,
  dirichlet_prior = 0.001,
  component_truncation = 7,
  regime_truncation = 15,
  hyperprior_b = 20,
  model_params_save_every = 5,
  simulation_iter = NULL,
  T2_window_size = 3,
  determining_p_cutoff = FALSE,
  prob_cutoff = 0.5,
  model_log_type = "NoModelSpecified",
  regime_selection_multiplicative_prior = 2,
  split_selection_multiplicative_prior = 2,
  is_initial_fit = TRUE,
  verbose = FALSE
)

Arguments

data_woTimeValues

matrix. Raw data matrix without datetime stamps.

time_of_observations

vector. Datetime stamps for every data instance in data_woTimeValues.

time_points

vector. Time points that mark each 'day' of time. Range should include every datetime in time_of_observations.

variable_names

vector. Vector of names of columnsof data_woTimeValues.

not.cont

vector. Indicator variable as to which columns are discrete.

iterations

integer. Number of MCMC samples to take (including burn-in).

burnin

integer. Number of burn-in samples. iterations > burnin necessarily.

lower_bounds

vector. Lower bounds for each data column.

upper_bounds

vector. Upper bounds for each data column.

ordinal_indicators

vector. Discrete values, one for each column, indicating which variables are ordinal.

list_of_ordinal_levels

vector. Discrete values, one for each column, indicating which variables are part of the same ordinal group.

categorical_indicators

vector. Each nominal d categorical variable must be broken down into d different indicator variables. This vector marks which variables are such indicators.

previous_states

vector. Starting regime vector, if known, of the same length as the number of 'days' in time_points.

previous_model_fits

rlist. Starting parameter fits corresponding to regime vector previous_states.

linger_parameter

float. Prior parameter for Markov chain probability matrix. Larger = less likely to change states.

move_parameter

float. Prior parameter for Markov chain probability matrix. Larger = more likely to change states.

g.prior

float in (0,1). Prior probability on edge inclusion for graph structure G.

set_G

matrix. Starting graph structure, if known.

wishart_df_initial

integer (>= 3). Starting DF for G-Wishart prior.

lambda

float. Parameter for NI-G-W prior, controls affect of precision sample on the center sample.

g_sampling_distribution

matrix. Prior probability on edge inclusion if not uniform across G.

n.cores

integer. Number of cores available for parallelization.

scaleMatrix

matrix. Parameter for NI-G-W prior.

allow_for_mixture_models

logical. Whether or not method should fix mixture distributions to regimes.

dirichlet_prior

float. Parameter for the dirichlet process for fitting components in the mixture model.

component_truncation

integer. Maximum component allowed. Should be sufficiently large.

regime_truncation

integer. Maximum regime allowed. Should be sufficiently large.

hyperprior_b

integer. Hyperprior on Wishart distribution fit to the scaleMatrix.

model_params_save_every

integer. How frequently to save model fits for the fault detection method.

simulation_iter

integer. Used for simulation studies. Deprecated value at package launch.

T2_window_size

integer. Length of sliding window for Hotelling T2 pre-step. Used when an initial value for previous_states is not provided.

determining_p_cutoff

logical. Method for estimating the probability cutoff on the posterior distribution for determining change-points. Deprecated at package launch date.

prob_cutoff

float. Changepoints are determined (for fault detection process) if posterior probability exceeds this value.

model_log_type

character vector. The type of log (used to distinguish logfiles).

regime_selection_multiplicative_prior

float. Must be >=1. Gives additional probability to the most recent day for the selection of a new split point.

split_selection_multiplicative_prior

float.

is_initial_fit

logical. True when there is no previously fit bayesWatch object fed through the algorithm..

verbose

logical. Prints verbose model output for debugging when TRUE. It is highly recommended that you pipe this to a text file.

Value

bayesWatch object. A model fit for the analysis of posterior change-points and fault detection.

Examples

library(bayesWatch)
data("full_data")
data("day_of_observations")
data("day_dts")

x       = bayeswatch(full_data, day_of_observations, day_dts,
                   iterations = 500, g.prior = 1, linger_parameter = 20, n.cores=3,
                   wishart_df_initial = 3, hyperprior_b = 3, lambda = 5)

print(x)
plot(x)
detect_faults(x)