- Updated LANDMark version

- TreeOrdination can now take advantage of LANDMark's proximity measures (both using terminal nodes as features or all nodes in the decision path as features)
- Data is cast into the np.float32 dtype for the CLRClosureTransformer

TreeOrdination/TreeOrdination.py

@@ -5,6 +5,9 @@
 from sklearn.base import ClassifierMixin, BaseEstimator, clone
 from sklearn.decomposition import PCA
 
+from scipy.sparse import csr_array
+from scipy.sparse import hstack as sp_hstack
+
 from umap import UMAP
 
 from LANDMark import LANDMarkClassifier
@@ -61,6 +64,7 @@ def __init__(
         feature_names,
         resampler=None,
         metric="hamming",
+        prox_method = "terminal",
         supervised_clf=ExtraTreesClassifier(1024),
         proxy_model = ExtraTreesRegressor(1024),
         landmark_model = LANDMarkClassifier(160, use_nnet=False, n_jobs = 8),
@@ -76,6 +80,7 @@ def __init__(
         self.resampler = resampler
 
         self.metric = metric
+        self.prox_method = prox_method
 
         self.supervised_clf = supervised_clf
         self.proxy_model = proxy_model
@@ -117,20 +122,22 @@ def get_initial_embedding(self, X):
 
             # Get proximity
             X_trf = resampler.transform(X)
-            self.LM_emb.append(model.proximity(X_trf))
+
+            # Update Overall Proximity
+            if i > 0:
+                self.LM_emb = sp_hstack((self.LM_emb, model.proximity(X_trf, self.prox_method)))
+            else:
+                self.LM_emb = model.proximity(X_trf, self.prox_method)
 
             # Save the resampler
             self.transformers.append(resampler)
 
-        # Get Overall Proximity
-        self.LM_emb = np.hstack(self.LM_emb)
-
         # Get Embeddings
         self.UMAP_trf = UMAP(
             n_neighbors=self.n_neighbors,
             n_components=15,
             min_dist=self.min_dist,
-            metric=self.metric,
+            metric="hamming",
             densmap=False,
         ).fit(self.LM_emb)
 
@@ -248,7 +255,7 @@ def plot_projection(self, X, y, ax_1=0, ax_2=1, use_approx=True, trf_type = "PCA
 
     def predict_proba(self, X):
 
-        tree_emb = self.emb_transform(X, "LM")
+        tree_emb = self.emb_transform(X, "UMAP")
 
         P = self.p_model.predict_proba(tree_emb)
 
@@ -281,11 +288,10 @@ def emb_transform(self, X, trf_type = "PCA"):
                 transformer = self.transformers[i]
 
                 # Get proximity
-                proximity = model.proximity(transformer.transform(X))
-
-                tree_emb.append(proximity)
-
-            tree_emb = np.hstack(tree_emb)
+                if i != 0:
+                    tree_emb = sp_hstack((tree_emb, model.proximity(transformer.transform(X), self.prox_method)))
+                else:
+                    tree_emb = model.proximity(transformer.transform(X), self.prox_method)
 
             if trf_type == "LM":
                 return tree_emb

TreeOrdination/transformers_treeord.py

@@ -16,18 +16,18 @@ def __init__(self, do_clr=False, delta=None):
     def fit_transform(self, X, y=None, **kwargs):
 
         if self.do_clr:
-            return clr(multiplicative_replacement(closure(X), delta=self.delta))
+            return clr(multiplicative_replacement(closure(X), delta=self.delta)).astype(np.float32)
 
         else:
-            return closure(X)
+            return closure(X).astype(np.float32)
 
     def transform(self, X, y=None, **kwargs):
 
         if self.do_clr:
-            return clr(multiplicative_replacement(closure(X), delta=self.delta))
+            return clr(multiplicative_replacement(closure(X), delta=self.delta)).astype(np.float32)
 
         else:
-            return closure(X)
+            return closure(X).astype(np.float32)
 
 
 class ResampleRandomizeTransform(BaseEstimator, TransformerMixin):

environment.yml

@@ -4,7 +4,7 @@ dependencies:
     - python >=3.8
     - pip
     - pip:
-        - LANDMarkClassifier >= 2.0.0
+        - LANDMarkClassifier >= 2.1.0
         - numpy == 1.23.5
         - scikit-learn >= 1.1.2
         - scikit-bio >= 0.5.8

pyproject.toml

@@ -4,7 +4,7 @@ requires = ["hatchling"]
 
 [project]
 name = "TreeOrdination"
-version = "1.3.2"
+version = "1.3.4"
 authors = [
     {name = "Josip Rudar", email = "[email protected]"},
     {name = "G. Brian Golding"},
@@ -40,7 +40,7 @@ dependencies = [
     "umap-learn >= 0.5.3",
     "seaborn",
     "shap >= 0.40.0",
-    "LANDMarkClassifier >= 2.0.4"
+    "LANDMarkClassifier >= 2.1.0"
 ]
 
 [tool.hatch.metadata]