lgb.interprete.R

#' Compute feature contribution of prediction
#'
#' Computes feature contribution components of rawscore prediction.
#'
#' @param model object of class \code{lgb.Booster}.
#' @param data a matrix object or a dgCMatrix object.
#' @param idxset an integer vector of indices of rows needed.
#' @param num_iteration number of iteration want to predict with, NULL or <= 0 means use best iteration.
#'
#' @return
#'
#' For regression, binary classification and lambdarank model, a \code{list} of \code{data.table} with the following columns:
#' \itemize{
#'   \item \code{Feature} Feature names in the model.
#'   \item \code{Contribution} The total contribution of this feature's splits.
#' }
#' For multiclass classification, a \code{list} of \code{data.table} with the Feature column and Contribution columns to each class.
#'
#' @examples
#' Sigmoid <- function(x) 1 / (1 + exp(-x))
#' Logit <- function(x) log(x / (1 - x))
#' data(agaricus.train, package = "lightgbm")
#' train <- agaricus.train
#' dtrain <- lgb.Dataset(train$data, label = train$label)
#' setinfo(dtrain, "init_score", rep(Logit(mean(train$label)), length(train$label)))
#' data(agaricus.test, package = "lightgbm")
#' test <- agaricus.test
#'
#' params <- list(
#'     objective = "binary"
#'     , learning_rate = 0.01
#'     , num_leaves = 63
#'     , max_depth = -1
#'     , min_data_in_leaf = 1
#'     , min_sum_hessian_in_leaf = 1
#' )
#' model <- lgb.train(params, dtrain, 10)
#'
#' tree_interpretation <- lgb.interprete(model, test$data, 1:5)
#'
#' @importFrom data.table as.data.table
#' @export
lgb.interprete <- function(model,
                           data,
                           idxset,
                           num_iteration = NULL) {

  # Get tree model
  tree_dt <- lgb.model.dt.tree(model, num_iteration)

  # Check number of classes
  num_class <- model$.__enclos_env__$private$num_class

  # Get vector list
  tree_interpretation_dt_list <- vector(mode = "list", length = length(idxset))

  # Get parsed predictions of data
  pred_mat <- t(
    model$predict(
      data[idxset, , drop = FALSE]
      , num_iteration = num_iteration
      , predleaf = TRUE
    )
  )
  leaf_index_dt <- data.table::as.data.table(pred_mat)
  leaf_index_mat_list <- lapply(
    X = leaf_index_dt
    , FUN = function(x) matrix(x, ncol = num_class, byrow = TRUE)
  )

  # Get list of trees
  tree_index_mat_list <- lapply(leaf_index_mat_list,
                                FUN = function(x) matrix(seq_len(length(x)) - 1, ncol = num_class, byrow = TRUE))

  # Sequence over idxset
  for (i in seq_along(idxset)) {
    tree_interpretation_dt_list[[i]] <- single.row.interprete(tree_dt, num_class, tree_index_mat_list[[i]], leaf_index_mat_list[[i]])
  }

  # Return interpretation list
  return(tree_interpretation_dt_list)

}

#' @importFrom data.table data.table
single.tree.interprete <- function(tree_dt,
                                   tree_id,
                                   leaf_id) {

  # Match tree id
  single_tree_dt <- tree_dt[tree_index == tree_id, ]

  # Get leaves
  leaf_dt <- single_tree_dt[leaf_index == leaf_id, .(leaf_index, leaf_parent, leaf_value)]

  # Get nodes
  node_dt <- single_tree_dt[!is.na(split_index), .(split_index, split_feature, node_parent, internal_value)]

  # Prepare sequences
  feature_seq <- character(0)
  value_seq <- numeric(0)

  # Get to root from leaf
  leaf_to_root <- function(parent_id, current_value) {

    # Store value
    value_seq <<- c(current_value, value_seq)

    # Check for null parent id
    if (!is.na(parent_id)) {

      # Not null means existing node
      this_node <- node_dt[split_index == parent_id, ]
      feature_seq <<- c(this_node[["split_feature"]], feature_seq)
      leaf_to_root(this_node[["node_parent"]], this_node[["internal_value"]])

    }

  }

  # Perform leaf to root conversion
  leaf_to_root(leaf_dt[["leaf_parent"]], leaf_dt[["leaf_value"]])

  # Return formatted data.table
  data.table::data.table(Feature = feature_seq, Contribution = diff.default(value_seq))

}

#' @importFrom data.table := rbindlist setorder
multiple.tree.interprete <- function(tree_dt,
                                     tree_index,
                                     leaf_index) {

  # Apply each trees
  interp_dt <- data.table::rbindlist(
    l = mapply(
      FUN = single.tree.interprete
      , tree_id = tree_index
      , leaf_id = leaf_index
      , MoreArgs = list(
        tree_dt = tree_dt
      )
      , SIMPLIFY = FALSE
      , USE.NAMES = TRUE
    )
    , use.names = TRUE
  )

  interp_dt <- interp_dt[, .(Contribution = sum(Contribution)), by = "Feature"]

  # Sort features in descending order by contribution
  interp_dt[, abs_contribution := abs(Contribution)]
  data.table::setorder(
    x = interp_dt
    , -abs_contribution
  )

  # Drop absolute value of contribution (only needed for sorting)
  interp_dt[, abs_contribution := NULL]

  return(interp_dt)

}

#' @importFrom data.table set setnames
single.row.interprete <- function(tree_dt, num_class, tree_index_mat, leaf_index_mat) {

  # Prepare vector list
  tree_interpretation <- vector(mode = "list", length = num_class)

  # Loop throughout each class
  for (i in seq_len(num_class)) {

    next_interp_dt <- multiple.tree.interprete(
      tree_dt = tree_dt
      , tree_index = tree_index_mat[,i]
      , leaf_index = leaf_index_mat[,i]
    )

    if (num_class > 1){
      data.table::setnames(
        next_interp_dt
        , old = "Contribution"
        , new = paste("Class", i - 1)
      )
    }

    tree_interpretation[[i]] <- next_interp_dt

  }

  # Check for numbe rof classes larger than 1
  if (num_class == 1) {

    # First interpretation element
    tree_interpretation_dt <- tree_interpretation[[1]]

  } else {

    # Full interpretation elements
    tree_interpretation_dt <- Reduce(f = function(x, y) merge(x, y, by = "Feature", all = TRUE),
                                     x = tree_interpretation)

    # Loop throughout each tree
    for (j in 2:ncol(tree_interpretation_dt)) {

      data.table::set(tree_interpretation_dt,
                      i = which(is.na(tree_interpretation_dt[[j]])),
                      j = j,
                      value = 0)

    }

  }

  # Return interpretation tree
  return(tree_interpretation_dt)
}