Merge pull request #31 from cosanlab/mult_update

use binary mask to improve NMF_mult fitting

neighbors/_fit.py

@@ -118,8 +118,9 @@ def sgd(
 
 
 @nb.njit(cache=True, nogil=True)
-def mult(X, W, H, data_range, eps, tol, n_iterations, verbose):
-    """Lee & Seung (2001) multiplicative update rule"""
+def mult(X, M, W, H, data_range, eps, tol, n_iterations, verbose):
+    """Lee & Seung (2001) multiplicative update rule extended with
+    Zhu (2016) binary 'weight' matrix to handle missing data."""
 
     last_e = 0
     error_history = np.zeros((n_iterations))
@@ -143,20 +144,23 @@ def mult(X, W, H, data_range, eps, tol, n_iterations, verbose):
                 tol,
             )
 
-        # Update H
-        numer = W.T @ X
-        denom = W.T @ W @ H + eps
+        # The np.multiply's below have the effect of only using observed (non-missing)
+        # ratings when performing the factor matrix updates
+
+        # Update H (factor x item)
+        numer = W.T @ np.multiply(M, X)
+        denom = W.T @ np.multiply(M, W @ H) + eps
         H *= numer
         H /= denom
 
-        # Update W
-        numer = X @ H.T
-        denom = W @ H @ H.T + eps
+        # Update W (user x factor)
+        numer = np.multiply(M, X) @ H.T
+        denom = np.multiply(M, W @ H) @ H.T + eps
         W *= numer
         W /= denom
 
         # Make prediction and get error
-        errors = X - W @ H
+        errors = np.multiply(M, X) - np.multiply(M, W @ H)
         rmse = np.sqrt(np.mean(np.power(errors, 2)))
 
         # Normalize current error with respect to max of dataset

neighbors/models.py

@@ -351,15 +351,19 @@ def fit(
         # Whereas in SGD we explity pass in indices of training data for fitting, here we set testing indices to 0 so they have no impact on the multiplicative update. See Zhu, 2016 for more details: https://arxiv.org/pdf/1612.06037.pdf
         self.dilate_mask(n_samples=dilate_by_nsamples)
 
-        # fillna(0) is equivalent to hadamard (element-wise) product with a binary mask
-        X = self.masked_data.fillna(0).to_numpy()
+        # Generate a binary mask matrix for observed ratings
+        M = self.mask.to_numpy().astype(float)
+        # mult() will compute hadamard (element-wise) product of the binary mask
+        # and all data to 0 out NaNs, so we need all data here
+        X = self.data.to_numpy()
 
         # Run multiplicative updating
         # Silence numba warning until this issue gets fixed: https://github.com/numba/numba/issues/4585
         with warnings.catch_warnings():
             warnings.filterwarnings("ignore", category=NumbaPerformanceWarning)
             error_history, converged, n_iter, delta, norm_rmse, W, H = mult(
                 X,
+                M,
                 self.W,
                 self.H,
                 self.data_range,