Data

Get all bike rides from my private Github Repository. Select relevant columns and parse start_date_local into a timestamp (start_timestamp_local). For some rides there is no gear_id. This is because for some old rides, I didn’t maintain this information. For these cases add ‘Endurace’ as gear. Deselect the no longer needed gear_id and start_date_local columns.

bike_rides <- function() {
  pins::board_register_github(repo = "duju211/strava_data")

  df_act_raw <- pins::pin_get("strava_activities", board = "github")

  pins::board_disconnect("github")

  rel_vars <- c(
    "id", "distance", "elapsed_time", "moving_time", "total_elevation_gain",
    "start_date_local", "gear_id", "average_speed", "max_speed",
    "average_temp", "elev_high", "elev_low", "pr_count")

  df_act_bike <- df_act_raw %>%
    dplyr::filter(type == "Ride") %>%
    dplyr::select(dplyr::one_of(rel_vars)) %>%
    dplyr::mutate(
      start_timestamp_local = lubridate::ymd_hms(start_date_local),
      gear = dplyr::case_when(
        gear_id == "b6601295" ~ "Aeroad",
        gear_id == "b4797231" ~ "Endurace",
        TRUE ~ "Endurace")) %>%
    dplyr::select(-c(start_date_local, gear_id))
}

Save the output of the function in the df_bike_rides data frame and take a quick look at the scatter plot matrix produced with the GGally package:

vis_bike_rides_eda <- function(df_bike_rides) {
  df_bike_rides %>%
    dplyr::select(-c(id, pr_count)) %>%
    GGally::ggscatmat(color = "gear", alpha = 0.6) +
    ggplot2::theme(axis.text.x = ggplot2::element_text(angle = 90))
}

Modelling

Create resamples of the data set:

init_split <- rsample::initial_split(df_bike_rides)

df_train <- rsample::training(init_split)
df_test <- rsample::testing(init_split)

df_resamples <- rsample::bootstraps(df_train)

Define the random forest model. Set the mode of the model to ‘regression’ and define mtry and min_n as tuning parameters.

rand_forest_model <- function() {
  cores <- parallel::detectCores()

  rf_mod <- parsnip::rand_forest(trees = 1000, mtry = tune(),
                                 min_n = tune()) %>%
    parsnip::set_engine(
      "ranger", num.threads = cores, importance = "impurity") %>%
    parsnip::set_mode("regression")
}

Define the preprocessing steps for the random forest model:

rand_forest_rec <- function(df_train) {
  strava_rec <- recipes::recipe(average_speed ~ ., data = df_train) %>%
    recipes::update_role(id, new_role = "ID") %>%
    recipes::step_rm(start_timestamp_local) %>%
    recipes::step_medianimpute(average_temp, elev_high, elev_low)
}

Combine the model and the preprocessing definition into one ‘workflow’

rand_forest_workflow <- function(rf_rec, rf_model) {
  workflows::workflow() %>%
    workflows::add_recipe(rf_rec) %>%
    workflows::add_model(rf_model)
}

Tune the workflow:

rand_forest_tune <- function(rf_workflow, df_resamples) {
  tune::tune_grid(
    rf_workflow, df_resamples,
    control = tune::control_grid(save_pred = TRUE))
}

Take a look at the result of the tuning process:

plot_tuning_results <- function(tune_res) {
  tune_res %>%
    tune::collect_metrics() %>%
    tidyr::pivot_longer(
      cols = c(mtry, min_n), names_to = "parameter", values_to = "value") %>%
    ggplot2::ggplot(ggplot2::aes(x = value, y = mean, color = .metric)) +
    ggplot2::geom_point() +
    ggplot2::facet_wrap(~ parameter, scales = "free_x")
}

Select the best combination of tuning parameters and finalise the workflow with these parameters:

best_rmse <- tune::select_best(rf_tune_res, metric = "rmse")
final_rf <- tune::finalize_workflow(rf_workflow, best_rmse)

The final workflow looks like this:

## == Workflow ==========================================================================================================================================
## Preprocessor: Recipe
## Model: rand_forest()
## 
## -- Preprocessor --------------------------------------------------------------------------------------------------------------------------------------
## 2 Recipe Steps
## 
## * step_rm()
## * step_medianimpute()
## 
## -- Model ---------------------------------------------------------------------------------------------------------------------------------------------
## Random Forest Model Specification (regression)
## 
## Main Arguments:
##   mtry = 3
##   trees = 1000
##   min_n = 3
## 
## Engine-Specific Arguments:
##   num.threads = cores
##   importance = impurity
## 
## Computational engine: ranger

Visualise the variable importance:

vis_vip <- function(final_rf, df_train) {
  final_rf %>%
    parsnip::fit(data = df_train) %>%
    workflows::pull_workflow_fit() %>%
    vip::vip(geom = "point")
}

readd(gg_vip, path = path_to_drake)