Please no changes this time.

R/RLearner_classif_lightgbm.R

@@ -12,7 +12,7 @@ makeRLearner.classif.lightgbm = function() {
       makeIntegerLearnerParam("eval_freq", lower = 1, tunable = FALSE, requires = quote(verbose > 0)),
       makeUntypedLearnerParam("init_model"),
       makeIntegerLearnerParam("early_stopping_rounds", lower = 1),
-      makeDiscreteLearnerParam("boosting", values = c("dbdt", "dart", "rf")),
+      makeDiscreteLearnerParam("boosting", values = c("gbdt", "dart", "rf"), default = "gbdt", requires = quote(boosting != "rf" || bagging_freq > 0 && bagging_fraction < 1 && feature_fraction < 1)),
       makeNumericLearnerParam("learning_rate", lower = 0, upper = 1, default = 0.1),
       makeIntegerLearnerParam("num_leaves", lower = 1, default = 31),
       makeDiscreteLearnerParam("tree_learner", values = c("serial", "feature", "data", "voting"), default = "serial"),
@@ -57,9 +57,9 @@ trainLearner.classif.lightgbm = function(.learner, .task, .subset, .weights = NU
   pv$metric = coalesce(metric, "")
 
   if(nc == 2) {
-    pv$obj = "binary"
+    pv$objective = "binary"
   } else {
-    pv$obj = "multiclass"
+    pv$objective = "multiclass"
     pv$num_class = nc
   }
 

R/RLearner_regr_lightgbm.R

@@ -0,0 +1,68 @@
+#' @export
+makeRLearner.regr.lightgbm = function() {
+  makeRLearnerRegr(
+    cl = "regr.lightgbm",
+    package = "lightgbm",
+    par.set = makeParamSet(
+      makeUntypedLearnerParam("validation.data"),
+      makeIntegerLearnerParam("nrounds", lower = 1, default = 10),
+      makeDiscreteLearnerParam("metric", values = c("l1", "l2", "l2_root", "quantile", "mape", "huber", "fair")),
+      makeDiscreteLearnerParam("obj", values = c("regression_l2", "regression_l1", "huber", "fair", "poisson", "quantile", "mape", "gamma", "tweedie", default = "regression_l2")),
+      makeIntegerLearnerParam("verbose", lower = -1, upper = 1, tunable = FALSE),
+      makeLogicalLearnerParam("record", default = TRUE, tunable = FALSE),
+      makeIntegerLearnerParam("eval_freq", lower = 1, tunable = FALSE, requires = quote(verbose > 0)),
+      makeUntypedLearnerParam("init_model"),
+      makeIntegerLearnerParam("early_stopping_rounds", lower = 1),
+      makeDiscreteLearnerParam("boosting", values = c("gbdt", "dart", "rf"), default = "gbdt", requires = quote(boosting != "rf" || bagging_freq > 0 && bagging_fraction < 1 && feature_fraction < 1)),
+      makeNumericLearnerParam("learning_rate", lower = 0, upper = 1, default = 0.1),
+      makeIntegerLearnerParam("num_leaves", lower = 1, default = 31),
+      makeDiscreteLearnerParam("tree_learner", values = c("serial", "feature", "data", "voting"), default = "serial"),
+      makeIntegerLearnerParam("num_threads", lower = 1),
+      makeDiscreteLearnerParam("device", values = c("cpu", "gpu"), default = "cpu"),
+      makeIntegerLearnerParam("max_depth", lower = -1, default = -1),
+      makeIntegerLearnerParam("min_data_in_leaf", lower = 1, default = 20),
+      makeNumericLearnerParam("min_sum_hessian_in_leaf", lower = 0, default = 1e-3),
+      makeNumericLearnerParam("feature_fraction", lower = 0, upper = 1, default = 1),
+      makeNumericLearnerParam("bagging_fraction", lower = 0, upper = 1, default = 1),
+      makeIntegerLearnerParam("bagging_freq", lower = 0, default = 0),
+      makeNumericLearnerParam("lambda_l1", lower = 0, default = 0),
+      makeNumericLearnerParam("lambda_l2", lower = 0, default = 0),
+      makeNumericLearnerParam("min_split_gain", lower = 0, default = 0),
+      makeNumericLearnerParam("drop_rate", lower = 0, upper = 1, default = 0.1, requires = quote(boosting == "dart")),
+      makeNumericLearnerParam("skip_drop", lower = 0, upper = 1, default = 0.5, requires = quote(boosting == "dart")),
+      makeIntegerLearnerParam("max_drop", lower = 0, default = 50, requires = quote(boosting == "dart")),
+      makeLogicalLearnerParam("xgboost_dart_mode", default = FALSE),
+      makeIntegerLearnerParam("min_data_per_group", lower = 1, default = 100),
+      makeIntegerLearnerParam("max_cat_threshold", lower = 0, default = 32),
+      makeNumericLearnerParam("cat_l2", lower = 0, default = 10),
+      makeIntegerLearnerParam("max_cat_to_onehot", lower = 0, default = 4)
+      ),
+    properties = c("numerics", "weights", "featimp", "missings", "factors"),
+    name = "Light Gradient Boosting Machine",
+    short.name = "lightgbm",
+    note = ""
+    )
+}
+
+#' @export
+trainLearner.regr.lightgbm = function(.learner, .task, .subset, .weights = NULL, validation.data = NULL, metric, ...) {
+
+  pv = list(...)
+  train = getTaskData(.task, .subset, target.extra = TRUE)
+  feat.cols = colnames(train$data)[vlapply(train$data, is.factor)]
+  prep = lgb.prepare_rules(train$data)
+  pv$data = lgb.Dataset(data.matrix(prep$data), label = as.numeric(train$target) - 1, categorical_feature = feat.cols)
+  if (!is.null(validation.data))
+    pv$valids = list(test = lgb.Dataset.create.valid(pv$data, data.matrix(validation.data$data), label = as.numeric(validation.data$target) - 1))
+  pv$metric = coalesce(metric, "")
+
+  mod = do.call(lgb.train, pv)
+  return(list(mod = mod, rules = prep$rules))
+}
+
+#' @export
+predictLearner.regr.lightgbm = function(.learner, .model, .newdata, ...) {
+  m = .model$learner.model
+  .newdata = data.matrix(lgb.prepare_rules(.newdata, rules = m$rules)$data)
+  predict(m$mod, .newdata)
+ }