fix the scores method

Author: jbiggsets

factor_analyzer/factor_analyzer.py

@@ -226,7 +226,14 @@ class FactorAnalyzer:
     rotation_matrix : np.array
         The rotation matrix, if a rotation
         has been performed.
-
+    structure : np.array or None
+        The structure loading matrix.
+        This only exists if the rotation
+        is promax.
+    psi : np.array or None
+        The factor correlations
+        matrix. This only exists
+        if the rotation is oblique.
     Notes
     -----
     This code was partly derived from the excellent R package
@@ -270,6 +277,9 @@ def __init__(self,
         self.loadings = None
         self.rotation_matrix = None
 
+        self._scale_mean = None
+        self._scale_std = None
+
     @staticmethod
     def _fit_uls_objective(psi, corr_mtx, n_factors):
         """
@@ -588,7 +598,7 @@ def fit_factor_analysis(self,
                              'Check to make sure you do not have any '
                              'features with zero standard deviation.')
 
-        corr = corr.values
+        corr = corr.copy().values
 
         # if `use_smc` is True, get get squared multiple correlations
         # and use these as initial guesses for optimizer
@@ -647,7 +657,6 @@ def analyze(self,
                 normalize=True,
                 impute='median',
                 remove_non_numeric=True,
-                use_scaling=True,
                 use_corr_matrix=False,
                 **kwargs):
         """
@@ -706,11 +715,6 @@ def analyze(self,
             Remove any non-numeric data. If `use_corr_matrix` is True,
             no non-numeric data will be removed.
             Defaults to True.
-        use_scaling : bool, optional
-            Whether to scale the data by subtracting out the mean
-            and dividing by the standard deviation. If `use_corr_matrix`
-            is True, scaling will not be performed.
-            Defaults to True.
         use_corr_matrix : bool, optional
             Set to true if the `data` is the correlation
             matrix.
@@ -772,16 +776,11 @@ def analyze(self,
             self.corr = df.copy()
         else:
             self.corr = df.corr()
-
-        # scale the data, if it is not a correlation
-        # matrix and `use_scaling` is True
-        if use_scaling and not use_corr_matrix:
-            X = (df - df.mean(0)) / df.std(0)
-        else:
-            X = df.copy()
+            self._scale_mean = df.mean(0)
+            self._scale_std = df.std(0)
 
         # fit factor analysis model
-        loadings = self.fit_factor_analysis(X,
+        loadings = self.fit_factor_analysis(df.copy(),
                                             n_factors,
                                             use_smc,
                                             bounds,
@@ -883,7 +882,7 @@ def get_eigenvalues(self):
         """
         if (self.corr is not None and self.loadings is not None):
 
-            corr = self.corr.values
+            corr = self.corr.copy().values
 
             e_values, _ = sp.linalg.eigh(corr)
             e_values = pd.DataFrame(e_values[::-1],
@@ -1023,20 +1022,46 @@ def get_factor_variance(self):
 
             return variance_info
 
-    def get_scores(self, data):
+    def get_scores(self,
+                   data,
+                   scale_mean=None,
+                   scale_std=None):
         """
         Get the factor scores, given the data.
 
         Parameters
         ----------
         data : pd.DataFrame
             The data to calculate factor scores.
+        scale_mean : float or None
+            The mean of the original
+            data set used to fit the
+            factor model. If None, attempt
+            to retrieve the mean from the
+            original `analyze()` method,
+            if it was saved.
+            Defaults to None.
+        scale_std : float or None
+            The standard deviation of the original
+            data set used to fit the
+            factor model. If None, attempt
+            to retrieve the standard deviation from the
+            original `analyze()` method,
+            if it was saved.
+            Defaults to None.
 
         Returns
         -------
         scores : pd.DataFrame
             The factor scores.
 
+        Raises
+        ------
+        ValueError
+            If either scale_std or scale_mean
+            is None, and the original mean or standard
+            deviation were not saved during fitting.
+
         Examples
         --------
         >>> import pandas as pd
@@ -1055,11 +1080,38 @@ def get_scores(self, data):
         if self.loadings is not None:
 
             df = data.copy()
-            corr = data.corr()
+            corr = self.corr.copy()
+
+            error_msg = ('The `{}` argument is None, but no scaled {} '
+                         'was saved when fitting your original factor '
+                         'model. This most likely because you used a '
+                         'correlation matrix, rather than the full data '
+                         'set. Please either pass a value for `{}` '
+                         'or re-fit your model using the full data set.')
+
+            # if no scaled mean is passed, use the mean from the
+            # original fitting procedure; otherwise, raise an error
+            if scale_mean is None and self._scale_mean is not None:
+                scale_mean = self._scale_mean
+            elif scale_mean is None and self._scale_mean is None:
+                raise ValueError(error_msg.format('scale_mean',
+                                                  'mean',
+                                                  'scale_mean'))
+
+            # if no scaled std is passed, use the std from the
+            # original fitting procedure; otherwise, raise an error
+            if scale_std is None and self._scale_std is not None:
+                scale_std = self._scale_std
+            elif scale_std is None and self._scale_std is None:
+                raise ValueError(error_msg.format('scale_std',
+                                                  'standard deviation',
+                                                  'scale_std'))
 
             # scale the data
-            X = (df - df.mean(0)) / df.std(0)
+            X = (df - scale_mean) / scale_std
 
+            # use the structure matrix, if it exists;
+            # otherwise, just use the loadings matrix
             if self.structure is not None:
                 structure = self.structure
             else:

factor_analyzer/rotator.py

@@ -235,6 +235,10 @@ def oblique(self,
         rotation_mtx : np.array
             The rotation matrix
             (n_factors, n_factors)
+        psi : np.array or None
+            The factor correlations
+            matrix. This only exists
+            if the rotation is oblique.
         """
         df = loadings.copy()
 
@@ -526,6 +530,10 @@ def promax(self, loadings, normalize=False, power=4):
         rotation_mtx : np.array
             The rotation matrix
             (n_factors, n_factors)
+        psi : np.array or None
+            The factor correlations
+            matrix. This only exists
+            if the rotation is oblique.
         """
         df = loadings.copy()
 
@@ -620,6 +628,10 @@ def rotate(self, loadings, method='varimax', **kwargs):
         rotation_mtx : np.array
             The rotation matrix
             (n_factors, n_factors)
+        psi : np.array or None
+            The factor correlations
+            matrix. This only exists
+            if the rotation is oblique.
 
         Raises
         ------

factor_analyzer/test_utils.py

@@ -72,8 +72,12 @@ def calculate_py_output(test_name,
 
     if use_corr_matrix:
         X = data.corr()
+        scale_mean = data.mean(0)
+        scale_std = data.std(0)
     else:
         X = data.copy()
+        scale_mean = None
+        scale_std = None
 
     rotation = None if rotation == 'none' else rotation
     method = {'uls': 'minres'}.get(method, method)
@@ -89,7 +93,7 @@ def calculate_py_output(test_name,
             'loading': fa.loadings,
             'uniquenesses': fa.get_uniqueness(),
             'communalities': fa.get_communalities(),
-            'scores': fa.get_scores(data)}
+            'scores': fa.get_scores(data, scale_mean, scale_std)}
 
 
 def collect_r_output(test_name,

tests/test_factor_analyzer.py

@@ -201,12 +201,13 @@ def test_analyze_rotation_value_error(self):
     @raises(ValueError)
     def test_analyze_infinite(self):
 
-        data = pd.DataFrame({'A': [2, 4, 5, 6, 8, 9],
-                             'B': [4, 8, float('inf'), 10, 16, 18],
-                             'C': [6, 12, 15, 12, 26, 27]})
+        data = pd.DataFrame({'A': [1.0, 0.4, 0.5],
+                             'B': [0.4, 1.0, float('inf')],
+                             'C': [0.5, float('inf'), 1.0]},
+                            index=['A', 'B', 'C'])
 
         fa = FactorAnalyzer()
-        fa.analyze(data, 1, impute='drop')
+        fa.analyze(data, 1, impute='drop', use_corr_matrix=True)
 
     def test_remove_all_columns(self):
         # test that columns with string values are removed.