Switched inverse function to run pseudo-inverse, to account for correlated and 0 variance variables

Kruskals/kruskals.py

@@ -67,7 +67,7 @@ def driver_score(self, directional=False):
             self._driver_score = (pij_row_mean + pijm_row_mean) / ((ind_c - 1) + fact)
             if directional:
                 self._driver_score = self._driver_score * np.apply_along_axis(self.correlation_coef, 0, self._ndarr, self._arr)
-        return self._driver_score
+        return np.nan_to_num(self._driver_score)
 
     def percentage(self):
         """
@@ -94,7 +94,7 @@ def pcor_squared(ndarr):
         """
         Internal method to calculate the partial correlation squared
         """
-        icvx = np.linalg.inv(np.cov(ndarr))
+        icvx = np.linalg.pinv(np.cov(ndarr))
         return (icvx[0, 1] * icvx[0, 1]) / (icvx[0, 0] * icvx[1, 1])
 
     def percentage(self):

README.md

@@ -6,6 +6,10 @@ Kruskal's Driver Analysis (Not to be confused with his Distance Measure algorith
 
 This package provides a python implementation of [Kruskal's Algorithm](https://en.wikipedia.org/wiki/Kruskal%27s_algorithm)
 
+Caveats
+------------
+
+To calculate the inverse it uses the (Moore–Penrose pseudoinverse)[https://en.wikipedia.org/wiki/Moore%E2%80%93Penrose_pseudoinverse] which permits highly correlated independent variables to be passed as well as variables that have zero variance. It is up to the user of this library to ensure they are comfortable with this. N.B. if the normal matrix inversion would work, that is used, the psuedoinverse is only applied if the former fails.
 
 Installation
 ------------

circle.yml

@@ -9,5 +9,3 @@ dependencies:
 test:
   override:
     - detox
-    - py.test --cov=./
-    - codecov --token=88ffd259-37b7-4c79-89a2-b4a00b83b519

tests/test_kruskals.py

@@ -121,3 +121,26 @@ def test_that_direction_is_applied_on_directional_drivers_analysis():
     series = Kruskals.Kruskals(ndarr, arr).driver_score_to_series(True)
 
     assert (series.values < 0).any()
+
+def test_ability_to_handle_all_same_type():
+    """
+    Test to make sure that kruskals can handle data
+    when all the values for and independent set are 0
+    """
+    ndarr = np.array([
+      [10, 0, 3, 4, 5, 6],
+      [6, 0, 4, 3, 5, 1],
+      [1, 0, 9, 1, 5, 1],
+      [9, 0, 2, 2, 5, 2],
+      [3, 0, 3, 9, 5, 3],
+      [1, 0, 2, 9, 5, 4],
+      [1, 0, 2, 9, 5, 4],
+      [1, 0, 2, 9, 5, 4]
+    ])
+
+    arr = np.array([1, 2, 3, 4, 5, 6, 7, 8])
+
+    series = Kruskals.Kruskals(ndarr, arr).driver_score()
+
+    assert series[1] == 0.0
+    assert series[4] == 0.0

tox.ini

@@ -1,7 +1,12 @@
 [tox]
-skipsdist = True
+skipsdist = true
 envlist = py27,py34,py35,py36
 [testenv]
+# necessary to make cov find the .coverage file
+# see http://blog.ionelmc.ro/2014/05/25/python-packaging/
+usedevelop = true
+skip_install = true
 commands =
-  py.test
+  py.test --cov=./
+  codecov --token=88ffd259-37b7-4c79-89a2-b4a00b83b519
   publish: python setup.py sdist upload --sign -r pypi