diff --git a/implicit/evaluation.pyx b/implicit/evaluation.pyx
index f25d0da..e029520 100644
--- a/implicit/evaluation.pyx
+++ b/implicit/evaluation.pyx
@@ -187,16 +187,17 @@ cpdef leave_k_out_split(
 
     # get only users with n + 1 interactions
     candidate_mask = counts > K + 1
+    if sum(candidate_mask) == 0:
+        return ratings.tocsr(), csr_matrix(ratings.shape)
+    unique_candidate_users = unique_users[candidate_mask]
 
     # keep a given subset of users _only_ in the training set.
     if train_only_size > 0.0:
-        train_only_mask = ~np.isin(
-            unique_users, _choose(random_state, len(unique_users), train_only_size)
-        )
-        candidate_mask = train_only_mask & candidate_mask
+        adjusted_ratio = min(1, (1 - train_only_size) / (unique_candidate_users.shape[0] / (unique_users.shape[0] + 1)))
+        train_only_mask = _choose(random_state, len(unique_candidate_users), adjusted_ratio)
+        unique_candidate_users = unique_candidate_users[train_only_mask]
 
     # get unique users who appear in the test set
-    unique_candidate_users = unique_users[candidate_mask]
     full_candidate_mask = np.isin(users, unique_candidate_users)
 
     # get all users, items and ratings that match specified requirements to be