Merge branch 'main' of https://github.com/abhi-glitchhg/multicolinearity

README.md

@@ -3,7 +3,7 @@
 
 # corrfeatred
 
-reduce features using correlation matrix 
+select features using correlation matrix 
 
 ## Installation 
 
@@ -29,15 +29,30 @@ different_feature_set = reduce_features(correlation_matrix, threshold=0.8, polic
 ```
 
 
-## Workflow
+## Idea and workflow
 
 Currently there is only one function which takes correlation matrix and thresholds as input and then constructs a graph. 
 
-There after we find maximal cliques in the graph and our goal is to have at max one feature from each clique.
+
+We create a graph where each node is represents a feature, and edge represents collinearity between the features. Then maximal cliques present in the graph are calculated. 
+
+
+Each clique represents a cluster of features that are correlated with each other, and hence only one feature from this cluster is enough to represent whole cluster in the final feature sets. Hence, we can have multiple policies about how we want to choose the features (minimum number of features, maximum number of features etc).
+
+Our goal is to have at max one feature from each clique.
+
+And finally the feature set we get from this function will all have pairwise correlation less than the threshhold. 
 
 ![workflow](https://github.com/abhi-glitchhg/corrfeatred/assets/72816663/731c0be4-75a0-4355-b4aa-7682d7759d38)
 
 
 
 
 
+
+
+
+
+
+
+

corrfeatred/reduce.py

@@ -12,12 +12,12 @@ def reduce_features(correlation_matrix, threshold=0.75,policy='min', random_stat
 #     correlation_matrix : df.corr
 #     threshold: float
 #     method: whether we want minimum number of features or maximum number of features. NOTE: this is bit unstable and sometimes `min` policy has more features than `max`, this depends on how graph develops. 
-#     random_state: random state, use this to get different set of features for same correlation matrix.  
+#     random_seed: random state, use this to get different set of features for same correlation matrix.  
 #     """
 
-    if random_state!=None:
-        random_gen = random.Random(random_state)
-    corrmatrix = correlation_matrix.copy()
+    if random_seed!=None:
+        random_gen = random.Random(random_seed)
+    corrmatrix = correlation_matrix.abs().copy()
     inf_ = 4*(corrmatrix.shape[0]) + 10 # adding 10 for no reason, this could be any positive number;
     corr_matrix = corrmatrix > threshold
     assert policy in ('min', 'max'), "wrong input parameter"
@@ -50,7 +50,7 @@ def reduce_features(correlation_matrix, threshold=0.75,policy='min', random_stat
                 node_to_clique_list[key].append(idx) # node to group mapping
     while not np.all(mask):
 
-        if random_state==None:
+        if random_seed==None:
             top = method(node_to_clique_count.items(), key = lambda kv: kv[1])
         else:
             dict_items = list(node_to_clique_count.items())

pyproject.toml

@@ -5,7 +5,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "corrfeatred"
-version = "0.0.3.1"
+version = "0.0.3.2"
 authors = [
   { name="Abhijit Deo", email="[email protected]" },
 ]

tests/test_correlation.py

@@ -7,7 +7,7 @@ def test_hello():
 
 
 def test_correlation():
-    for i in (4,8,10,12,16,22,100,1000):
+    for i in (4,8,10,12,16,22,100,500):
         rand_array = np.random.uniform(0,1,i*i).reshape(i,i)
         corr_arr = np.tril(rand_array) + np.triu(rand_array.T,1)
         assert (corr_arr!=corr_arr.T).sum().sum() == 0