Fits a 2-component Gaussian Mixture Model (GMM) on transformed movement
features and decodes each row into inactive/active states using posterior
probabilities. Optional median or HMM smoothing can be applied to reduce label
flicker.
Usage
grz_classify_activity_gmm(
data,
groups = NULL,
step_col = "step_m",
turn_col = "turn_rad",
feature_set = c("adaptive", "legacy"),
adaptive_window_mins = "auto",
adaptive_window_mult = 4,
adaptive_window_min_mins = 30,
state_col = "activity_state_gmm",
component_col = "activity_component_gmm",
inactive_prob_col = "inactive_prob_gmm",
fit_max_rows = 200000L,
max_iter = 200L,
tol = 1e-05,
min_var = 1e-06,
smoothing = c("none", "median", "hmm"),
median_window_n = 5L,
hmm_self_transition = 0.98,
seed = 1,
verbose = TRUE,
return_class = c("data.frame", "data.table")
)Arguments
- data
Input GPS data.
- groups
Grouping columns used for track-wise decoding.
- step_col
Step-distance column (meters).
- turn_col
Turn-angle column (radians).
- feature_set
GMM feature set.
"adaptive"(default) augments step and turn with adaptive-window displacement features."legacy"uses only step and turn.- adaptive_window_mins
Adaptive feature window size in minutes. Use
"auto"(default) to scale window length from each track's sampling interval.- adaptive_window_mult
Multiplier applied to base sampling interval when
adaptive_window_mins = "auto".- adaptive_window_min_mins
Lower bound for auto window size (minutes).
- state_col
Output state column name.
- component_col
Output numeric component id column.
- inactive_prob_col
Output posterior probability column for inactive state.
- fit_max_rows
Maximum rows used to fit the GMM (sampled if needed).
- max_iter
Maximum EM iterations.
- tol
EM convergence tolerance on log-likelihood.
- min_var
Minimum per-component feature variance for numerical stability.
- smoothing
Posterior smoothing method:
"none","median", or"hmm".- median_window_n
Window size (number of rows) for median smoothing when
smoothing = "median".- hmm_self_transition
Self-transition probability used when
smoothing = "hmm".- seed
Random seed for reproducible fitting/subsampling.
- verbose
Logical; print summary output.
- return_class
Output class:
"data.frame"(default) or"data.table".