Add comment for creating additional synthetic dataset

README.md

@@ -1,5 +1,23 @@
-# Beyond Spectral Clustering
+# EECS-553-Machine Learning (Reproducibility Project)
+This repository is for reproducibility project from EECS553 (Machine Leaning)
+Course. We verified the paper "A Tighter Analysis of Spectral Clustering, and
+Beyond", published in ICML 2022.
+
+## Additional Test that we excecuted
+1. **Less-separated Synthetic Dataset**: run
+`python experiments.py complete`
+- 
+
+2. **Test on BSDS dataset with different standard deviation**: run
+`python experiments.py bsds`
+- 
+
+3. **Test on MNIST dataset with different number of eigenvector for embedding**:
+   run
 
+
+
+# Beyond Spectral Clustering
 This directory contains the code to reproduce the results in the paper "A Tighter Analysis of Spectral Clustering, and Beyond", published in 
 ICML 2022.
 

experiments.py

@@ -147,8 +147,12 @@ def __init__(self, k, n, prob_p, prob_q, queue, num_runs=1, use_grid=False,
         self.queue = queue
         self.num_runs = num_runs
         self.use_grid = use_grid
-        self.use_complete = use_complete
+
+        # [Dongmyeong Lee] Additianl condition to generate synthetic dataset,
+        #                  which is less-separated clusters
+        self.use_complete = use_complete 
         self.unequal_cluster = unequal_cluster
+
         self.d = self.k
         if use_grid:
             self.k = self.d * self.d
@@ -429,14 +433,20 @@ def main():
 
     if args.experiment == 'cycle':
         run_sbm_experiment(1000, 5, 0.01)
+    # [Dongmyeong Lee] Generate Unbalanced Clusters that form Cyclic Structure
     elif args.experiment == 'cycle_unequal':
         run_sbm_experiment([100, 120, 140, 160, 180, 200], 6, 0.01, unequal_cluster=True)
     elif args.experiment == 'grid':
         run_sbm_experiment(1000, 5, 0.01, use_grid=True)
+    # [Dongmyeong Lee] Generate Unbalanced Clusters that form Grid Structure
     elif args.experiment == 'grid_unequal':
         run_sbm_experiment([100 + 20 * i for i in range(4*4)], 4, 0.01, use_grid=True, unequal_cluster=True)
+    # [Dongmyeong Lee] Generate Less-separated Clusters that similar to Cyclic Structure
+    # p: probability of having edges between nodes in same cluster
+    # q: probability of having edges between nodes in adjacent clusters
+    # r: probability of having edges between nodes in non-adjacent clusters // Inside SbmCompleteDataset Class
     elif args.experiment == 'complete':
-        run_sbm_experiment(100, 5, 0.2, use_complete=True)
+        run_sbm_experiment(1000, 5, 0.01, use_complete=True)
     elif args.experiment == 'mnist':
         run_mnist_experiment()
     elif args.experiment == 'usps':

pysc/datasets.py

@@ -361,7 +361,8 @@ def __str__(self):
     def __repr__(self):
         return self.__str__()
 
-
+# [Dongmyeong Lee] New Dataset Class for generating less-separated Clusters
+#                  Manually changed r {0.1, 0.3, 0.5, 0.7, 0.9}
 class SbmCompleteDataset(Dataset):
 
     def __init__(self, *args, k=6, n=50, p=0.3, q=0.05, **kwargs):
@@ -375,8 +376,10 @@ def load_data(self, data_file):
     def load_graph(self, graph_file=None, graph_type="knn10"):
         # Generate the graph from the sbm
         logger.info(f"Generating {self} graph from sbm...")
+        r = self.q * 0.9
         prob_mat = self.p * sp.sparse.eye(self.k) + \
-                   self.q * sgtl.graph.complete_graph(self.k).adjacency_matrix()
+                   (self.q - r) * sgtl.graph.cycle_graph(self.k).adjacency_matrix() + \
+                   r * sgtl.graph.complete_graph(self.k).adjacency_matrix()
         self.graph = sgtl.random.sbm_equal_clusters(self.n * self.k, self.k,
                                                     prob_mat.toarray())
 
@@ -396,7 +399,8 @@ def __str__(self):
     def __repr__(self):
         return self.__str__()
 
-
+# [Dongmyeong Lee] New Synthetic Dataset to generate unbalanaced clusters that
+# form Cylcic Structure
 class SbmUnequalCycleDataset(Dataset):
 
     def __init__(self, *args, k=6, n=[10, 20, 30, 40, 50, 60], p=0.3, q=0.05, **kwargs):
@@ -431,6 +435,8 @@ def __repr__(self):
         return self.__str__()
 
 
+# [Dongmyeong Lee] New Synthetic Dataset to generate unbalanaced clusters that
+# form Grid Structure
 class SbmUnequalGridDataset(Dataset):
 
     def __init__(self, *args, d=3, n=[10, 20, 30, 40, 50, 60, 70, 80, 90], p=0.3, q=0.05, **kwargs):

results/sbm/complete_results.csv

@@ -0,0 +1 @@
+k, n, p, q, poverq, eigenvectors, conductance, rand