diff --git a/Orange/widgets/data/owfeatureconstructor.py b/Orange/widgets/data/owfeatureconstructor.py
index 72704e9727c..674f1d0fe8d 100644
--- a/Orange/widgets/data/owfeatureconstructor.py
+++ b/Orange/widgets/data/owfeatureconstructor.py
@@ -21,7 +21,7 @@
 from traceback import format_exception_only
 from collections import namedtuple, OrderedDict
 from itertools import chain, count, starmap
-from typing import List, Dict, Any, Mapping
+from typing import List, Dict, Any, Mapping, Optional
 
 import numpy as np
 
@@ -1021,6 +1021,7 @@ def bind_variable(descriptor, env, data, use_values):
 
     values = {}
     cast = None
+    dtype = object if isinstance(descriptor, StringDescriptor) else float
 
     if isinstance(descriptor, DiscreteDescriptor):
         if not descriptor.values:
@@ -1038,7 +1039,7 @@ def bind_variable(descriptor, env, data, use_values):
         cast = DateTimeCast()
 
     func = FeatureFunc(descriptor.expression, source_vars, values, cast,
-                       use_values=use_values)
+                       use_values=use_values, dtype=dtype)
     return descriptor, func
 
 
@@ -1216,7 +1217,10 @@ class FeatureFunc:
         A function for casting the expressions result to the appropriate
         type (e.g. string representation of date/time variables to floats)
     """
-    def __init__(self, expression, args, extra_env=None, cast=None, use_values=False):
+    dtype: Optional['DType'] = None
+
+    def __init__(self, expression, args, extra_env=None, cast=None, use_values=False,
+                 dtype=None):
         self.expression = expression
         self.args = args
         self.extra_env = dict(extra_env or {})
@@ -1225,6 +1229,7 @@ def __init__(self, expression, args, extra_env=None, cast=None, use_values=False
         self.cast = cast
         self.mask_exceptions = True
         self.use_values = use_values
+        self.dtype = dtype
 
     def __call__(self, table, *_):
         if isinstance(table, Table):
@@ -1252,7 +1257,7 @@ def __call_table(self, table):
             y = list(starmap(f, args))
         if self.cast is not None:
             y = self.cast(y)
-        return y
+        return np.asarray(y, dtype=self.dtype)
 
     def __call_instance(self, instance: Instance):
         table = Table.from_numpy(
@@ -1281,7 +1286,8 @@ def extract_column(self, table: Table, var: Variable):
 
     def __reduce__(self):
         return type(self), (self.expression, self.args,
-                            self.extra_env, self.cast, self.use_values)
+                            self.extra_env, self.cast, self.use_values,
+                            self.dtype)
 
     def __repr__(self):
         return "{0.__name__}{1!r}".format(*self.__reduce__())
diff --git a/Orange/widgets/data/tests/test_owfeatureconstructor.py b/Orange/widgets/data/tests/test_owfeatureconstructor.py
index 3de43bae642..237b4908ac7 100644
--- a/Orange/widgets/data/tests/test_owfeatureconstructor.py
+++ b/Orange/widgets/data/tests/test_owfeatureconstructor.py
@@ -8,6 +8,7 @@
 from unittest.mock import patch, Mock
 
 import numpy as np
+from scipy import sparse as sp
 
 from orangewidget.settings import Context
 
@@ -148,6 +149,20 @@ def test_unicode_normalization():
                                      construct_variables(desc, data)))
         np.testing.assert_equal(data.X, data.metas)
 
+    def test_transform_sparse(self):
+        domain = Domain([ContinuousVariable("A")])
+        desc = [
+            ContinuousDescriptor(name="X", expression="A", number_of_decimals=2)
+        ]
+        X = sp.csc_matrix(np.arange(5).reshape(5, 1))
+        data = Table.from_numpy(domain, X)
+        data_ = data.transform(Domain(data.domain.attributes,
+                                      [],
+                                      construct_variables(desc, data)))
+        np.testing.assert_equal(
+            data.get_column_view(0)[0], data_.get_column_view(0)[0]
+        )
+
 
 class TestTools(unittest.TestCase):
     def test_free_vars(self):
@@ -276,7 +291,7 @@ def test_reconstruct(self):
 
     def test_repr(self):
         self.assertEqual(repr(FeatureFunc("a + 1", [("a", 2)])),
-                         "FeatureFunc('a + 1', [('a', 2)], {}, None, False)")
+                         "FeatureFunc('a + 1', [('a', 2)], {}, None, False, None)")
 
     def test_call(self):
         iris = Table("iris")
@@ -291,7 +306,7 @@ def test_string_casting(self):
         f = FeatureFunc("name[0]",
                         [("name", zoo.domain["name"])])
         r = f(zoo)
-        self.assertEqual(r, [x[0] for x in zoo.metas[:, 0]])
+        self.assertEqual(list(r), [x[0] for x in zoo.metas[:, 0]])
         self.assertEqual(f(zoo[0]), str(zoo[0, "name"])[0])
 
     def test_missing_variable(self):
@@ -309,7 +324,7 @@ def test_time_str(self):
         data = Table.from_numpy(Domain([TimeVariable("T", have_date=True)]), [[0], [0]])
         f = FeatureFunc("str(T)", [("T", data.domain[0])])
         c = f(data)
-        self.assertEqual(c, ["1970-01-01", "1970-01-01"])
+        self.assertEqual(list(c), ["1970-01-01", "1970-01-01"])
 
     def test_invalid_expression_variable(self):
         iris = Table("iris")