diff --git a/R/PipeOpDecode.R b/R/PipeOpDecode.R
index 489e5e3eb..edf6f07c6 100644
--- a/R/PipeOpDecode.R
+++ b/R/PipeOpDecode.R
@@ -5,6 +5,7 @@
 #' @format [`R6Class`][R6::R6Class] object inheriting from [`PipeOpTaskPreprocSimple`]/[`PipeOpTaskPreproc`]/[`PipeOp`].
 #'
 #' @description
+#' Description
 #'
 #' @section Construction:
 #' ```
@@ -50,7 +51,7 @@ PipeOpDecode = R6Class("PipeOpDecode",
         group_pattern = p_uty(custom_check = check_string, tags = c("train", "predict"))
       )
       ps$values = list(treatment_encoding = FALSE, group_pattern = "^([^.]*)\\.")
-      super$initialize(id, param_set = ps, param_vals = param_vals, packages = "stats", tags = "encode", feature_types = c("factor", "ordered"))
+      super$initialize(id, param_set = ps, param_vals = param_vals, tags = "encode", feature_types = c("integer", "numeric"))
     }
   ),
   private = list(
@@ -62,42 +63,62 @@ PipeOpDecode = R6Class("PipeOpDecode",
       # If pattern == "", all columns are collapsed into one column
       if (pv$group_pattern == "") {
         return(list(colmaps = list(result = set_names(cols, cols))))
+        # should make sure that result is available (or name we chose instead)
       }
 
-      # Extract group names
+      # Drop columns that do contain group_pattern
+      # What about cols starting with .? -> probably let user exclude this by changing group_pattern are using affect_columns
+      cols = cols[grepl(pv$group_pattern, cols)]
+
+      # Extract factor names
       matches = regmatches(cols, regexec(pv$group_pattern, cols))
-      grps = unlist(map(matches, function(x) if (length(x)) x[[2]] else ""))
+      fcts = vapply(matches, function(x) x[[2]], character(1))
       # Extract level names
       lvls = set_names(gsub(pv$group_pattern, "", cols), cols)
 
-      # Drop entries for which no match to group_pattern was found
-      keep = fcts != ""
-      fcts = fcts[keep]
-      lvls = lvls[keep]
-
-      # add "" = "ref" if pv$treatment_encoding == TRUE
       # test that split is consistent for this use case
-      list(colmaps = split(lvls, fcts))
-    },
+      s = list(colmaps = split(lvls, fcts))
+
+      if (pv$treatment_encoding) {
+        # Set default name for reference level
+        ref_name = "ref"
+        counter = 1
+        while (ref_name %in% cols) {
+          ref_name = paste0("ref.", counter)
+          counter = counter + 1
+        }
+        # Append ref_name with empty name to all list entries
+        for (i in seq_along(s[["colmaps"]])) {
+          s[["colmaps"]][[i]][[length(s[["colmaps"]][[i]]) + 1]] = ref_name
+        }
+      }
 
-    # take maximum value, bc could be scaled
-    # treatment dass alles 0 ist, hard coden, referenzname als reference nennen (und ref.1 falls es die spalte schon gibt)
+      s
+    },
 
-    # decide when to assign "ref" (e.g. no unique maximum)
     .transform_dt = function(dt, levels) {
       colmaps = self$state$colmaps
 
       for (fct in names(colmaps)) {
         old_cols = names(colmaps[[fct]])
-        lvls = unname(colmaps[[fct]])
+        lvls = colmaps[[fct]]
+
+        # Do we check that which.max is unique?
+        # Generally what checks should we perform? e.g. that group_pattern contains a capturing group
 
         # Find the column with the maximal value for each row
         dt[, (fct) := old_cols[apply(.SD, 1, which.max)], .SDcols = old_cols]
         # Assign the corresponding value from the named vector to the new column
-        dt[, (fct) := lvls[get(fct)]]
+        # dt[, (fct) := lvls[get(fct)]]
+        dt[, (fct) := lvls[dt[[fct]]]]
+
+        # type conversion to factor?
+
         # Remove the old columns (can move this to outside the loop)
         dt[, (old_cols) := NULL]
       }
+
+      dt
     }
   )
 )