[Model] add dhn (#207)

* [model] add dhn

* Modify the dhn code and use MessagePassing

* Add dataset and test code

* Add dataset and test code

* modify the code and add test code

* update

* update

* update

---------

Co-authored-by: Guangyu Zhou <77875480+gyzhou2000@users.noreply.github.com>
Co-authored-by: gyzhou2000 <gyzhou2000@gmail.com>
Co-authored-by: gyzhou2000 <gyzhou@bupt.edu.cn>

Author: 3 people

docs/source/api/gammagl.utils.rst

@@ -27,4 +27,5 @@ gammagl.utils
     gammagl.utils.to_scipy_sparse_matrix
     gammagl.utils.read_embeddings
     gammagl.utils.homophily
-    gammagl.utils.get_train_val_test_split
+    gammagl.utils.get_train_val_test_split
+    gammagl.utils.find_all_simple_paths

examples/dhn/README.md

@@ -0,0 +1,17 @@
+# Distance encoding based Heterogeneous graph neural Network (DHN)
+- Paper link: [https://ieeexplore.ieee.org/document/10209229](https://ieeexplore.ieee.org/document/10209229)
+- Author's code repo: [https://github.com/BUPT-GAMMA/HDE](https://github.com/BUPT-GAMMA/HDE)
+
+## Dataset Statics
+| Dataset  | # Nodes | # Edges |
+|----------|---------|---------|
+| acm      | 3908    | 4500    |
+
+## Results
+```bash
+TL_BACKEND="torch" python dhn_trainer.py --test_ratio 0.3 --one_hot True --k_hop 2 --num_neighbor 5 --batch_size 32 --lr 0.001 --n_epoch 100 --drop_rate 0.01 --dataset 'acm'
+```
+
+| Dataset  | Paper(AUC) | Our(th)(AUC)     |
+| -------- | ----- | ----------- | 
+| acm      | 95.07 |  95.54±0.18 | 

examples/dhn/dhn_trainer.py

@@ -0,0 +1,339 @@
+import os
+# os.environ['CUDA_VISIBLE_DEVICES'] = '0'
+# os.environ['TL_BACKEND'] = 'torch'
+os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2' 
+# 0:Output all; 1:Filter out INFO; 2:Filter out INFO and WARNING; 3:Filter out INFO, WARNING, and ERROR
+
+import argparse
+import random
+import numpy as np
+import tensorlayerx as tlx
+from tensorlayerx.model import TrainOneStep
+from sklearn.metrics import roc_auc_score
+from gammagl.models import DHNModel
+from gammagl.datasets import ACM4DHN
+from gammagl.utils import k_hop_subgraph, find_all_simple_paths
+
+
+type2idx = {
+    'M': 0,
+    'A': 1,
+    # 'C': 2,
+    # 'T': 3
+}
+
+
+def dist_encoder(src, dest, G, k_hop):
+    if (G.size(1) == 0):
+        paths = []
+    else:
+        paths = find_all_simple_paths(G, src, dest, k_hop + 2)
+
+    node_type = len(type2idx)
+    cnt = [k_hop + 1] * node_type  # Default truncation for max_spd exceeded
+    for path in paths:
+        res = [0] * node_type
+        for i in path:
+            if i >= 0:
+                res[type2idx['M']] += 1
+            else:
+                res[type2idx['A']] += 1
+
+        for k in range(node_type):
+            cnt[k] = min(cnt[k], res[k])
+
+    # Generate one-hot encoding
+    if args.one_hot:
+        one_hot_list = [np.eye(k_hop + 2, dtype=np.float64)[cnt[i]]
+                        for i in range(node_type)]
+        return np.concatenate(one_hot_list)
+    return cnt
+
+
+def type_encoder(node):
+    node_type = len(type2idx)
+    res = [0] * node_type
+    if node.item() >= 0:
+        res[type2idx['M']] = 1.0
+    else:
+        res[type2idx['A']] = 1.0
+    return res
+
+
+mini_batch = []
+fea_batch = []
+
+
+def gen_fea_batch(G, root, fea_dict, k_hop):
+    fea_batch = []
+    mini_batch.append(root)
+
+    a = [0] * (k_hop + 2) * 4 + type_encoder(root)
+
+    node_type = len(type2idx)
+    num_fea = (k_hop + 2) * 4 + node_type
+    fea_batch.append(np.asarray(a,
+                                dtype=np.float32
+                                ).reshape(-1, num_fea)
+                     )
+
+    # 1-order neighbor sampling
+    ns_1 = []
+    src, dst = G
+    for node in mini_batch[-1]:
+        if node.item() >= 0:
+            neighbors_mask = src == node
+        else:
+            neighbors_mask = dst == node
+        neighbors = list(tlx.convert_to_numpy(dst[neighbors_mask]))
+        neighbors.append(node.item())
+        random_choice_list = np.random.choice(neighbors, args.num_neighbor, replace=True)
+        ns_1.append(random_choice_list.tolist())
+    ns_1 = tlx.convert_to_tensor(ns_1)
+    mini_batch.append(ns_1[0])
+
+    de_1 = [
+        np.concatenate([fea_dict[ns_1[0][i].item()], np.asarray(type_encoder(ns_1[0][i]))], axis=0)
+        for i in range(0, ns_1[0].shape[0])
+    ]
+
+    fea_batch.append(np.asarray(de_1,
+                                dtype=np.float32).reshape(1, -1)
+                     )
+
+    # 2-order neighbor sampling
+    ns_2 = []
+    for node in mini_batch[-1]:
+        if node.item() >= 0:
+            neighbors_mask = src == node
+        else:
+            neighbors_mask = dst == node
+        neighbors = list(tlx.convert_to_numpy(dst[neighbors_mask]))
+        neighbors.append(node.item())
+        random_choice_list = np.random.choice(neighbors, args.num_neighbor, replace=True)
+        ns_2.append(random_choice_list.tolist())
+    ns_2 = tlx.convert_to_tensor(ns_2)
+
+    de_2 = []
+    for i in range(len(ns_2)):
+        tmp = []
+        for j in range(len(ns_2[0])):
+            tmp.append(
+                np.concatenate([fea_dict[ns_2[i][j].item()], np.asarray(type_encoder(ns_2[i][j]))], axis=0)
+            )
+        de_2.append(tmp)
+
+    fea_batch.append(np.asarray(de_2,
+                                dtype=np.float32).reshape(1, -1)
+                     )
+
+    return np.concatenate(fea_batch, axis=1)
+
+
+def subgraph_sampling_with_DE_node_pair(G, node_pair, k_hop=2):
+    [A, B] = node_pair
+
+    edge_index = tlx.concat([G['M', 'MA', 'A'].edge_index, reversed(G['M', 'MA', 'A'].edge_index)], axis=1)
+
+    # Find k-hop subgraphs of A and B
+    sub_G_for_AB = k_hop_subgraph([A, B], k_hop, edge_index)
+
+    # Remove edges using Boolean indexes
+    # Note: Just remove the edges, the points remain
+    edge_index_np = tlx.convert_to_numpy(sub_G_for_AB[1])
+    remove_indices = tlx.convert_to_tensor([
+        ((edge_index_np[0, i] == A) & (edge_index_np[1, i] == B)) | (
+                (edge_index_np[0, i] == B) & (edge_index_np[1, i] == A))
+        for i in range(sub_G_for_AB[1].shape[1])
+    ])
+    remove_indices = tlx.convert_to_numpy(remove_indices)
+    sub_G_index = sub_G_for_AB[1][:, ~remove_indices]
+
+    sub_G_nodes = set(np.unique(tlx.convert_to_numpy(sub_G_for_AB[0]))) | set(
+        np.unique(tlx.convert_to_numpy(sub_G_for_AB[1])))  # Gets the points in the graph
+    sub_G_nodes = tlx.convert_to_tensor(list(sub_G_nodes))
+
+    # Distance from all points in the subgraph to the node pair
+    SPD_based_on_node_pair = {}
+    for node in sub_G_nodes:
+        tmpA = dist_encoder(A, node, sub_G_index, k_hop)
+        tmpB = dist_encoder(B, node, sub_G_index, k_hop)
+
+        SPD_based_on_node_pair[node.item()] = np.concatenate([tmpA, tmpB], axis=0)
+
+    A_fea_batch = gen_fea_batch(sub_G_index, A,
+                                SPD_based_on_node_pair, k_hop)
+    B_fea_batch = gen_fea_batch(sub_G_index, B,
+                                SPD_based_on_node_pair, k_hop)
+
+    return A_fea_batch, B_fea_batch
+
+
+def batch_data(G, batch_size):
+    edge_index = G['M', 'MA', 'A'].edge_index
+    nodes = set(tlx.convert_to_tensor(np.unique(tlx.convert_to_numpy(edge_index[0])))) | set(
+        tlx.convert_to_tensor(np.unique(tlx.convert_to_numpy(edge_index[1]))))
+
+    nodes_list = []
+    for node in nodes:
+        nodes_list.append(node.item())
+
+    num_batch = int(len(edge_index[0]) / batch_size)
+
+    # Shuffle the order of the edges
+    edge_index_np = tlx.convert_to_numpy(edge_index)
+    permutation = np.random.permutation(edge_index_np.shape[1])  # Generate a randomly arranged index
+    edge_index_np = edge_index_np[:, permutation]  # Use this permutation index to scramble edge_index
+    edge_index = tlx.convert_to_tensor(edge_index_np)
+
+    for idx in range(num_batch):
+        batch_edge = edge_index[:, idx * batch_size:(idx + 1) * batch_size]  # Take out batch_size edges
+        batch_label = [1.0] * batch_size
+
+        batch_A_fea = []
+        batch_B_fea = []
+        batch_x = []
+        batch_y = []
+
+        i = 0
+        for by in batch_label:
+            bx = batch_edge[:, i:i + 1]
+
+            # Positive sample
+            posA, posB = subgraph_sampling_with_DE_node_pair(G, bx, k_hop=args.k_hop)
+            batch_A_fea.append(posA)
+            batch_B_fea.append(posB)
+            batch_y.append(np.asarray(by, dtype=np.float32))
+
+            # Negative sample
+            neg_tmpB_id = random.choice(nodes_list)
+            node_pair = tlx.convert_to_tensor([[bx[0].item()], [neg_tmpB_id]])
+
+            negA, negB = subgraph_sampling_with_DE_node_pair(G, node_pair, k_hop=args.k_hop)
+            batch_A_fea.append(negA)
+            batch_B_fea.append(negB)
+            batch_y.append(np.asarray(0.0, dtype=np.float32))
+
+        yield np.asarray(np.squeeze(batch_A_fea)), np.asarray(np.squeeze(batch_B_fea)), np.asarray(
+            batch_y).reshape(batch_size * 2, 1)
+
+
+class Loss(tlx.model.WithLoss):
+    def __init__(self, net, loss_fn):
+        super(Loss, self).__init__(backbone=net, loss_fn=loss_fn)
+
+    def forward(self, data, y):
+        logits = self.backbone_network(data['n1'], data['n2'], data['label'])
+        y = tlx.convert_to_tensor(y)
+        loss = self._loss_fn(logits, y)
+        return loss
+
+
+class AUCMetric:
+    def __init__(self):
+        self.true_labels = []
+        self.predicted_scores = []
+
+    def update_state(self, y_true, y_pred):
+        self.true_labels.extend(y_true)
+        self.predicted_scores.extend(y_pred)
+
+    def result(self):
+        auc = roc_auc_score(self.true_labels, self.predicted_scores)
+        return auc
+
+
+def main(args):
+    if str.lower(args.dataset) not in ['acm']:
+        raise ValueError('Unknown dataset: {}'.format(args.dataset))
+    if str.lower(args.dataset) == 'acm':
+        data = ACM4DHN(root=args.dataset_path, test_ratio=args.test_ratio)
+
+    graph = data[0]
+
+    G_train = graph['train']
+    G_val = graph['val']
+    G_test = graph['test']
+
+    node_type = len(type2idx)
+    num_fea = (args.k_hop + 2) * 4 + node_type
+
+    model = DHNModel(num_fea, args.batch_size, args.num_neighbor, name="DHN")
+
+    optim = tlx.optimizers.Adam(lr=args.lr, weight_decay=args.drop_rate)
+    train_weights = model.trainable_weights
+
+    net_with_loss = Loss(model, loss_fn=tlx.losses.sigmoid_cross_entropy)
+    net_with_train = TrainOneStep(net_with_loss, optim, train_weights)
+
+    tra_auc_metric = AUCMetric()
+    val_auc_metric = AUCMetric()
+    test_auc_metric = AUCMetric()
+
+    best_val_auc = 0
+    for epoch in range(args.n_epoch):
+
+        # train
+        model.set_train()
+        tra_batch_A_fea, tra_batch_B_fea, tra_batch_y = batch_data(G_train, args.batch_size).__next__()
+        tra_out = model(tra_batch_A_fea, tra_batch_B_fea, tra_batch_y)
+
+        data = {
+            "n1": tra_batch_A_fea,
+            "n2": tra_batch_B_fea,
+            "label": tra_batch_y
+        }
+
+        tra_loss = net_with_train(data, tra_batch_y)
+        tra_auc_metric.update_state(y_true=tra_batch_y, y_pred=tlx.convert_to_numpy(tlx.sigmoid(tra_out)))
+        tra_auc = tra_auc_metric.result()
+
+        # val
+        model.set_eval()
+        val_batch_A_fea, val_batch_B_fea, val_batch_y = batch_data(G_val, args.batch_size).__next__()
+        val_out = model(val_batch_A_fea, val_batch_B_fea, val_batch_y)
+
+        val_auc_metric.update_state(y_true=val_batch_y, y_pred=tlx.convert_to_numpy(tlx.sigmoid(val_out)))
+        val_auc = val_auc_metric.result()
+
+        print("Epoch [{:0>3d}] ".format(epoch+1)\
+              + "  train loss: {:.4f}".format(tra_loss.item())\
+              + "  val auc: {:.4f}".format(val_auc))
+
+        if val_auc > best_val_auc:
+            best_val_auc = val_auc
+            model.save_weights(args.best_model_path+model.name+".npz", format='npz_dict')
+
+    model.load_weights(args.best_model_path+model.name+".npz", format='npz_dict')
+    # test
+    test_batch_A_fea, test_batch_B_fea, test_batch_y = batch_data(G_test, args.batch_size).__next__()
+    test_out = model(test_batch_A_fea, test_batch_B_fea, test_batch_y)
+
+    test_auc_metric.update_state(y_true=test_batch_y, y_pred=tlx.convert_to_numpy(tlx.sigmoid(test_out)))
+    test_auc = test_auc_metric.result()
+    print("Test auc:  {:.4f}".format(test_auc))
+
+
+if __name__ == '__main__':
+    # parameters setting
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--test_ratio", type=float, default=0.3, help="ratio of dividing the data set")
+    parser.add_argument("--one_hot", type=bool, default=True, help="use one-hot encoding")
+    parser.add_argument("--k_hop", type=int, default=2, help="hops of the generated subgraph")
+    parser.add_argument("--num_neighbor", type=int, default=5, help="neighbor sample number")
+    parser.add_argument("--batch_size", type=int, default=32, help="batch size")
+    parser.add_argument("--lr", type=float, default=0.001, help="learning rate")
+    parser.add_argument("--n_epoch", type=int, default=100, help="number of epoch")
+    parser.add_argument("--drop_rate", type=float, default=0.01, help="drop_rate")
+    parser.add_argument('--dataset', type=str, default='acm', help='dataset')
+    parser.add_argument("--dataset_path", type=str, default=r"", help='dataset_path')
+    parser.add_argument("--best_model_path", type=str, default=r'./', help="path to save best model")
+    parser.add_argument("--gpu", type=int, default=-1)
+
+    args = parser.parse_args()
+    if args.gpu >= 0:
+        tlx.set_device("GPU", args.gpu)
+    else:
+        tlx.set_device("CPU")
+
+    main(args)

gammagl/datasets/__init__.py

@@ -23,6 +23,7 @@
 from .facebook import FacebookPagePage
 from .acm4heco import ACM4HeCo
 from .yelp import Yelp
+from .acm4dhn import ACM4DHN
 
 __all__ = [
     'ACM4HeCo',
@@ -48,7 +49,8 @@
     'MoleculeNet',
     'FacebookPagePage',
     'NGSIM_US_101',
-    'Yelp'
+    'Yelp',
+    'ACM4DHN'
 ]
 
 classes = __all__