feat: data preparation related scripts added

data_loader.py

@@ -0,0 +1,120 @@
+import scipy.io as sio
+import numpy as np
+import torch
+
+class DataLoader:
+    """Data Loader class"""
+    def __init__(self, S_train, S_val, S_test, data_dir, mat_file_path):
+        self.S_train = S_train
+        self.S_val = S_val
+        self.S_test = S_test
+        self.data_root_dir = data_dir
+        self.mat_file_path = mat_file_path
+
+        self.saved_content = sio.loadmat(self.data_root_dir + self.mat_file_path)
+        self.T = np.max(self.saved_content["A_labels_subs"][0, :]) + 1
+        print("Number of total graphs: {}".format(self.T))
+        self.N = max(np.max(self.saved_content["A_labels_subs"][1, :]),
+                     np.max(self.saved_content["A_labels_subs"][2, :])) + 1
+        print("Number of nodes in each graph: {}".format(self.N))
+
+        self.A_size = torch.Size([self.T, self.N, self.N])  # size of the adjacency matrix
+        self.C_size = torch.Size([self.T, self.N, self.N])  # similar to the adjacency matrix, C tensor has TxNxN shape
+        # labels of the edges
+        self.A_labels = torch.sparse.FloatTensor(torch.tensor(self.saved_content["A_labels_subs"], dtype=torch.long),
+                                                 torch.squeeze(torch.tensor(self.saved_content["A_labels_vals"])),
+                                                 self.A_size).coalesce()
+
+        # Laplacian Transformed of adjacency matrix
+        self.C = torch.sparse.FloatTensor(torch.tensor(self.saved_content["C_subs"], dtype=torch.long),
+                                          torch.squeeze(torch.tensor(self.saved_content["C_vals"])),
+                                          self.C_size).coalesce()
+
+        # adjacency matrix
+        self.A = torch.sparse.FloatTensor(self.A_labels._indices(),
+                                          torch.ones(self.A_labels._values().shape),
+                                          self.A_size).coalesce()
+        # create node features
+        self.X = self.create_node_features()
+
+    def split_data(self):
+        C_train = []
+        for j in range(self.S_train):
+            idx = self.C._indices()[0] == j
+            C_train.append(torch.sparse.FloatTensor(self.C._indices()[1:3, idx],
+                                                    self.C._values()[idx]))
+
+        C_val = []
+        for j in range(self.S_train, self.S_train + self.S_val):
+            idx = self.C._indices()[0] == j
+            C_val.append(torch.sparse.FloatTensor(self.C._indices()[1:3, idx],
+                                                    self.C._values()[idx]))
+
+        C_test = []
+        for j in range(self.S_train+self.S_test, self.S_train + self.S_val + self.S_test):
+            idx = self.C._indices()[0] == j
+            C_test.append(torch.sparse.FloatTensor(self.C._indices()[1:3, idx],
+                                                  self.C._values()[idx]))
+
+        C = {'C_train': C_train,
+                 'C_val': C_val,
+                 'C_test': C_test}
+
+        X_train = self.X[0:self.S_train].double()
+        X_val = self.X[self.S_train:self.S_train + self.S_val].double()
+        X_test = self.X[self.S_train + self.S_val:].double()
+
+        data = {'X_train' : X_train,
+                'X_val': X_val,
+                'X_test': X_test}
+
+
+        return data, C
+
+    def get_edges_and_labels(self):
+        # training
+        subs_train = self.A_labels._indices()[0] < self.S_train
+        edges_train = self.A_labels._indices()[:, subs_train]
+        labels_train = torch.sign(self.A_labels._values()[subs_train])
+        target_train = (labels_train != -1).long()    # element = 0 if class = -1; and 1 if class is 0 or +1
+
+        # validation
+        subs_val = (self.A_labels._indices()[0] >= self.S_train) & (self.A_labels._indices()[0] < self.S_train + self.S_val)
+        edges_val = self.A_labels._indices()[:, subs_val]
+        edges_val[0] -= self.S_train
+        labels_val = torch.sign(self.A_labels._values()[subs_val])
+        target_val = (labels_val != -1).long()
+
+        # Testing
+        subs_test = (self.A_labels._indices()[0] >= self.S_train + self.S_val)
+        edges_test = self.A_labels._indices()[:, subs_test]
+        edges_test[0] -= (self.S_train + self.S_val)
+        labels_test = torch.sign(self.A_labels._values()[subs_test])
+        target_test = (labels_test != -1).long()
+
+        targets = {'target_train': target_train,
+                   'target_val': target_val,
+                   'target_test': target_test}
+
+        edges = {'edges_train': edges_train,
+                 'edges_val': edges_val,
+                 'edges_test': edges_test}
+
+        return targets, edges
+
+    def create_node_features(self):
+        X = torch.zeros(self.A.shape[0], self.A.shape[1], 2)
+        X[:, :, 0] = torch.sparse.sum(self.A, 1).to_dense()   # number of outgoing edges
+        X[:, :, 1] = torch.sparse.sum(self.A, 2).to_dense()  # number of in coming edges
+
+        return X
+
+    def load_data(self):
+        print("Loading the data...")
+        data, C = self.split_data()
+        targets, edges = self.get_edges_and_labels()
+        print("======================")
+        return data, C, targets, edges
+
+
+

read_data.py

@@ -0,0 +1,291 @@
+import numpy as np
+import math
+import torch
+import scipy.io as sio
+import argparse
+
+parser = argparse.ArgumentParser()
+parser.add_argument('--data_path', type=str, help='path to the root of data directory')
+parser.add_argument('--dataset', type=str, help='dataset name')
+parser.add_argument('--time_partitioning', type=list, help='time partitioning for train/test/val split')
+
+args = parser.parse_args()
+    
+#Settings
+edge_life = True
+edge_life_window = 10
+no_diag = 20
+make_symmetric = True
+print(args.dataset)
+
+def print_tensor(A, str):
+    print('------------------------')
+    print(str)
+    print(A)
+    print(torch.sum(A._values()))
+    print('------------------------')
+
+if args.dataset == 'Bitcoin Alpha':
+    data = np.loadtxt(args.data_path, delimiter=',')
+    save_file_location = 'data/Bitcoin_Alpha/'
+    save_file_name = 'saved_content_bitcoin_alpha.mat'
+    time_delta = 60*60*24*14 # 2 weeks
+    time_slices = args.time_partitioning
+    no_train_samples, no_val_samples, no_test_samples = time_slices[0], time_slices[1], time_slices[2]
+
+if args.dataset == 'Bitcoin OTC':
+    data = np.loadtxt(args.data_path, delimiter=',')
+    save_file_location = 'data/Bitcoin_OTC/'
+    save_file_name = 'saved_content_bitcoin_otc.mat'
+    time_delta = 60*60*24*14 # 2 weeks
+    time_slices = args.time_partitioning
+    no_train_samples, no_val_samples, no_test_samples = time_slices[0], time_slices[1], time_slices[2]
+
+if args.dataset == 'Reddit':
+    data = np.loadtxt(args.data_path)
+    save_file_location = 'data/reddit/'
+    save_file_name = 'saved_content_reddit.mat'
+    time_delta = 60 * 60 * 24 * 14  # 2 weeks
+    time_slices = args.time_partitioning
+    no_train_samples, no_val_samples, no_test_samples = time_slices[0], time_slices[1], time_slices[2]
+
+elif args.dataset == 'Chess':
+    data = np.loadtxt(args.data_path, delimiter=',', skiprows=1)
+    #data = readmatrix('./data/chess/out.chess.csv')
+    save_file_location = '/home/shivmaran/Desktop/Tensor-GCN/data/chess/'
+    save_file_name = 'saved_content_python_chess.mat'
+    time_delta = 60*60*24*31 # 31 days
+    time_slices = args.time_partitioning
+    no_train_samples, no_val_samples, no_test_samples = time_slices[0], time_slices[1], time_slices[2]
+else:
+    print('Invalid dataset')
+    #exit
+
+data = torch.tensor(data)
+# Create full tensor
+if args.dataset == 'Chess':
+    dates = np.unique(data[:,3])
+    no_time_slices = len(dates)
+else:
+    no_time_slices = math.floor((max(data[:,3]) - min(data[:,3]))/time_delta)
+
+N = int(max(max(data[:,0]), max(data[:,1])))
+T = int(no_train_samples)
+TT = int(no_time_slices)
+
+#Create M
+M = np.zeros((T,T))
+for i in range(no_diag):
+    A = M[i:, :T-i]
+    np.fill_diagonal(A, 1)
+L = np.sum(M, axis=1)
+M = M/L[:,None]
+M = torch.tensor(M) 
+
+#Create A and A_labels
+if not args.dataset == 'Chess':
+    data = data[data[:,3] < min(data[:,3])+TT*time_delta]
+    start_time = min(data[:,3]);
+
+tensor_idx = torch.zeros([data.size()[0], 3], dtype=torch.long)
+tensor_val = torch.ones([data.size()[0]], dtype=torch.double)
+tensor_labels = torch.zeros([data.size()[0]], dtype=torch.double)
+
+for t in range(TT):
+    if args.dataset == 'Chess':
+        idx = data[:,3] == dates[t]
+    else:
+        end_time = start_time + time_delta
+        idx = (data[:, 3] >= start_time) & (data[:, 3] < end_time)
+        start_time = end_time
+    
+    tensor_idx[idx, 1:3] = (data[idx, 0:2] - 1).type('torch.LongTensor')
+    tensor_idx[idx, 0] = t
+    tensor_labels[idx] = data[idx, 2].type('torch.DoubleTensor')
+
+A =  torch.sparse.DoubleTensor(tensor_idx.transpose(1,0), tensor_val, torch.Size([TT, N, N])).coalesce()
+A_labels = torch.sparse.DoubleTensor(tensor_idx.transpose(1,0), tensor_labels, torch.Size([TT, N, N])).coalesce()
+
+def func_make_symmetric(sparse_tensor, N, TT):  
+    count = 0
+    tensor_idx = torch.LongTensor([])
+    tensor_val = torch.DoubleTensor([]).unsqueeze(1)
+    A_idx = sparse_tensor._indices()
+    A_val = sparse_tensor._values()
+    for j in range(TT):
+        idx = A_idx[0] == j
+        mat = torch.sparse.DoubleTensor(A_idx[1:3,idx], A_val[idx], torch.Size([N,N]))
+        mat_t = mat.transpose(1,0)
+        sym_mat = mat + mat_t
+        sym_mat = sym_mat/2
+        count = count + sym_mat._nnz()
+        vertices = torch.tensor(sym_mat._indices())
+        time = torch.ones(sym_mat._nnz(), dtype=torch.long)* j
+        time = time.unsqueeze(0)
+        full = torch.cat((time,vertices),0)
+        tensor_idx = torch.cat((tensor_idx,full),1)
+        tensor_val = torch.cat((tensor_val, sym_mat._values().unsqueeze(1)),0)        
+    tensor_val.squeeze_(1)
+    A =  torch.sparse.DoubleTensor(tensor_idx, tensor_val, torch.Size([TT, N, N])).coalesce()
+    return A
+
+if make_symmetric:
+    B = func_make_symmetric(A, N, TT)
+else:
+    B = A
+
+def func_edge_life(A, N, TT):
+    A_new = A.clone()
+    A_new._values()[:] = 0
+    idx = A._indices()[0] == 0
+    for t in range(TT):
+        idx =  (A._indices()[0] >= max(0, t-edge_life_window+1)) & (A._indices()[0] <= t)  
+        block = torch.sparse.DoubleTensor(A._indices()[0:3,idx], A._values()[idx], torch.Size([TT, N, N]))
+        block._indices()[0] = t
+        A_new = A_new + block
+    return A_new.coalesce()
+
+if edge_life:
+    B = func_edge_life(B,N,TT)
+
+def func_laplacian_transformation(B, N, TT):
+    vertices = torch.LongTensor([range(N), range(N)])
+    tensor_idx = torch.LongTensor([])
+    tensor_val = torch.DoubleTensor([]).unsqueeze(1)
+    for j in range(TT):
+        time = torch.ones(N, dtype=torch.long) * j
+        time = time.unsqueeze(0)
+        full = torch.cat((time,vertices),0)
+        tensor_idx = torch.cat((tensor_idx,full),1)
+        val = torch.ones(N, dtype=torch.double)
+        tensor_val = torch.cat((tensor_val, val.unsqueeze(1)),0) 
+    tensor_val.squeeze_(1)
+    I = torch.sparse.DoubleTensor(tensor_idx, tensor_val , torch.Size([TT,N,N]))
+    C = B + I
+    tensor_idx = torch.LongTensor([])
+    tensor_val = torch.DoubleTensor([]).unsqueeze(1)
+    for j in range(TT):
+        idx = C._indices()[0] == j
+        mat = torch.sparse.DoubleTensor(C._indices()[1:3,idx], C._values()[idx], torch.Size([N,N]))
+        vec = torch.ones([N,1], dtype=torch.double)
+        degree =  1/torch.sqrt(torch.sparse.mm(mat, vec))        
+        index = torch.LongTensor(C._indices()[0:3,idx].size())
+        values = torch.DoubleTensor(C._values()[idx].size())
+        index[0] = j
+        index[1:3] = mat._indices()
+        values = mat._values()
+        count = 0
+        for i,j in index[1:3].transpose(1,0):
+            values[count] = values[count] * degree[i] * degree[j]
+            count = count + 1
+        tensor_idx = torch.cat((tensor_idx,index), 1)
+        tensor_val = torch.cat((tensor_val,values.unsqueeze(1)),0)
+    tensor_val.squeeze_(1)
+    C = torch.sparse.DoubleTensor(tensor_idx, tensor_val , torch.Size([TT,N,N]))
+    return C.coalesce()
+
+C = func_laplacian_transformation(B, N, TT)
+Ct = C.clone().coalesce() 
+if TT < (T + no_val_samples + no_test_samples):
+    TTT= (T + no_val_samples + no_test_samples)
+    Ct = torch.sparse.DoubleTensor(Ct._indices(), Ct._values() , torch.Size([TTT,N,N])).coalesce()
+else:
+    TTT = TT
+
+def func_create_sparse(A, N, TTT, T, start, end):
+    assert (end-start) == T
+    idx = (A._indices()[0] >= start) & (A._indices()[0] < end)        
+    index = torch.LongTensor(A._indices()[0:3,idx].size())
+    values = torch.DoubleTensor(A._values()[idx].size())    
+    index[0:3] = A._indices()[0:3,idx]
+    index[0] = index[0] - start
+    values = A._values()[idx]
+    sub = torch.sparse.DoubleTensor(index, values , torch.Size([T,N,N]))
+    return sub.coalesce()
+
+
+C_train = func_create_sparse(Ct, N, TTT, T, 0, T)#There is a bug in matlab Bitcoin Alpha
+C_val = func_create_sparse(Ct, N, TTT, T, no_val_samples, T+no_val_samples)
+C_test = func_create_sparse(Ct, N, TTT, T, no_val_samples+no_test_samples, TTT)
+
+def to_sparse(x):
+    x_typename = torch.typename(x).split('.')[-1]
+    sparse_tensortype = getattr(torch.sparse, x_typename)
+
+    indices = torch.nonzero(x)
+    if len(indices.shape) == 0:  # if all elements are zeros
+        return sparse_tensortype(*x.shape)
+    indices = indices.t()
+    values = x[tuple(indices[i] for i in range(indices.shape[0]))]
+    return sparse_tensortype(indices, values, x.size())
+
+dense = torch.randn(3,3)
+dense[[0,0,1], [1,2,0]] = 0 # make sparse
+
+def func_MProduct(C, M):
+    assert C.size()[0] == M.size()[0]
+    Tr = C.size()[0]
+    N = C.size()[1]
+    C_new = torch.sparse.DoubleTensor(C.size())
+    #C_new = C.clone()
+    for j in range(Tr):
+        idx = C._indices()[0] == j
+        mat = torch.sparse.DoubleTensor(C._indices()[1:3,idx], C._values()[idx], torch.Size([N,N]))
+        tensor_idx = torch.zeros([3, mat._nnz()], dtype=torch.long)
+        tensor_val = torch.zeros([mat._nnz()], dtype=torch.double)
+        tensor_idx[1:3] = mat._indices()[0:2]
+        indices = torch.nonzero(M[:,j])
+        assert indices.size()[0] <= no_diag
+        for i in range(indices.size()[0]):
+            tensor_idx[0] = indices[i]
+            tensor_val = M[indices[i], j] * mat._values()
+            C_new = C_new + torch.sparse.DoubleTensor(tensor_idx, tensor_val , C.size())
+        C_new.coalesce()                      
+    return C_new.coalesce()  
+        
+Ct_train = func_MProduct(C_train, M)#There is a bug in matlab Bitcoin Alpha
+Ct_val = func_MProduct(C_val, M)
+Ct_test = func_MProduct(C_test, M)
+
+A_subs = A._indices()
+A_vals = A._values()
+A_labels_subs = A_labels._indices()
+A_labels_vals = A_labels._values()
+C_subs = C._indices()
+C_vals = C.values()
+C_train_subs = C_train._indices()
+C_train_vals = C_train.values()
+C_val_subs = C_val._indices()
+C_val_vals = C_val.values()
+C_test_subs = C_test._indices()
+C_test_vals = C_test.values()
+Ct_train_subs = Ct_train._indices()
+Ct_train_vals = Ct_train.values()
+Ct_val_subs = Ct_val._indices()
+Ct_val_vals = Ct_val.values()
+Ct_test_subs = Ct_test._indices()
+Ct_test_vals = Ct_test.values()
+
+sio.savemat(save_file_location + save_file_name, {
+        'tensor_idx': np.array(tensor_idx),
+        'tensor_labels': np.array(tensor_labels),
+        'A_labels_subs': np.array(A_labels_subs),
+        'A_labels_vals': np.array(A_labels_vals),
+        'A_subs': np.array(A_subs),
+        'A_vals': np.array(A_vals),
+        'C_subs': np.array(C_subs),
+        'C_vals': np.array(C_vals),
+        'C_train_subs': np.array(C_train_subs),
+        'C_train_vals': np.array(C_train_vals),
+        'C_val_subs': np.array(C_val_subs),
+        'C_val_vals': np.array(C_val_vals),
+        'C_test_subs': np.array(C_test_subs),
+        'C_test_vals': np.array(C_test_vals),
+        'Ct_train_subs': np.array(Ct_train_subs),
+        'Ct_train_vals': np.array(Ct_train_vals),
+        'Ct_val_subs': np.array(Ct_val_subs),
+        'Ct_val_vals': np.array(Ct_val_vals),
+        'Ct_test_subs': np.array(Ct_test_subs),
+        'Ct_test_vals': np.array(Ct_test_vals),
+        'M': np.array(M)
+    })

run.sh

@@ -0,0 +1,6 @@
+#PBS –N job1
+#PBS –m ae
+#PBS –M [email protected]
+#PBS –l nodes=1:ppn=5
+cd ./
+python -nodisplay -nodesktop -nojvm <./read_data.py > result.txt

utils.py

@@ -0,0 +1,39 @@
+import torch
+import pickle
+import os
+
+def f1_score(y_pred, y_true):
+    tp = torch.sum((y_pred == 0) & (y_true == 0), dtype=torch.float64)
+    fp = torch.sum((y_pred == 0) & (y_true != 0), dtype=torch.float64)
+    fn = torch.sum((y_pred != 0) & (y_true == 0), dtype=torch.float64)
+    precision = tp / (tp + fp)
+    recall = tp / (tp + fn)
+    f1 = 2 * (precision * recall) / (precision + recall)
+
+    return f1, precision, recall
+
+if __name__ == "__main__":
+
+    max_f1 = 0
+    fname = None
+    epoch = None
+    each_split = []
+
+    for bin_fname in os.listdir('results_edge_classification_bitcoin_otc/'):
+        bin_file = open("results_edge_classification_bitcoin_otc/" + bin_fname, "rb")
+        data = pickle.load(bin_file)
+        assert data.shape[1] == 12
+
+        for i in range(len(data)):
+            if data[i,2] == 1 or data[i,4] == 1 or data[i,6] == 1:
+                continue
+            mean_f1 = (data[i, 2] + data[i, 4] + data[i, 6] ) / 3
+            if mean_f1 > max_f1:
+                max_f1 = mean_f1
+                fname = bin_fname
+                epoch = i
+                each_split = [data[i, 2], data[i, 4], data[i, 6]]
+
+    print(fname, epoch, each_split,  str(max_f1))
+
+