GraphSAGE 代码解析 - minibatch.py

class EdgeMinibatchIterator

    """ This minibatch iterator iterates over batches of sampled edges or
    random pairs of co-occuring edges.

    G -- networkx graph
    id2idx -- dict mapping node ids to index in feature tensor
    placeholders -- tensorflow placeholders object
    context_pairs -- if not none, then a list of co-occuring node pairs (from random walks)
    batch_size -- size of the minibatches
    max_degree -- maximum size of the downsampled adjacency lists
    n2v_retrain -- signals that the iterator is being used to add new embeddings to a n2v model
    fixed_n2v -- signals that the iterator is being used to retrain n2v with only existing nodes as context
    """

def __init__(self, G, id2idx, placeholders, context_pairs=None, batch_size=100, max_degree=25,

n2v_retrain=False, fixed_n2v=False, **kwargs) 中具体介绍以下：

1 self.nodes = np.random.permutation(G.nodes())
2 # 函数shuffle与permutation都是对原来的数组进行重新洗牌,即随机打乱原来的元素顺序
3 # shuffle直接在原来的数组上进行操作，改变原来数组的顺序，无返回值
4 # permutation不直接在原来的数组上进行操作，而是返回一个新的打乱顺序的数组，并不改变原来的数组。

1 self.adj, self.deg = self.construct_adj()

这里重点看construct_adj()函数。

 def construct_adj(self):
         adj = len(self.id2idx) * \
             np.ones((len(self.id2idx) + 1, self.max_degree))
         # 该矩阵记录训练数据中各节点的邻居节点的编号
         # 采样只取max_degree个邻居节点，采样方法见下
         # 同样进行了行数加一操作

         deg = np.zeros((len(self.id2idx),))
         # 该矩阵记录了每个节点的度数

         for nodeid in self.G.nodes():
             if self.G.node[nodeid]['test'] or self.G.node[nodeid]['val']:
                 continue
             neighbors = np.array([self.id2idx[neighbor]
                                   for neighbor in self.G.neighbors(nodeid)
                                   if (not self.G[nodeid][neighbor]['train_removed'])])
             # Graph.neighbors() Return a list of the nodes connected to the node n.
             # 在选取邻居节点时进行了筛选，对于G.neighbors(nodeid) 点node的邻居，
             # 只取该node与neighbor相连的边的train_removed = False的neighbor
             # 也就是只取不是val, test的节点。
             # neighbors得到了邻居节点编号数列。

             deg[self.id2idx[nodeid]] = len(neighbors)
             # deg各位取值为该位对应nodeid的节点的度数，
             # 也即经过上面筛选后得到的邻居数

             if len(neighbors) == 0:
                 continue
             if len(neighbors) > self.max_degree:
                 neighbors = np.random.choice(
                     neighbors, self.max_degree, replace=False)
             # range: neighbors; size = max_degree; replace: replace the origin matrix or not
             # np.random.choice为选取size大小的数列

             elif len(neighbors) < self.max_degree:
                 neighbors = np.random.choice(
                     neighbors, self.max_degree, replace=True)
             # 经过choice随机选取，得到了固定大小max_degree = 25的直接相连的邻居数列

             adj[self.id2idx[nodeid], :] = neighbors
            # 把该node的邻居数列，赋值给adj矩阵中对应nodeid位的向量。
         return adj, deg

在construct_test_adj()  函数中，与上不同之处在于，可以直接得到邻居而无需根据val/test/train_removed筛选.

 neighbors = np.array([self.id2idx[neighbor]
                           for neighbor in self.G.neighbors(nodeid)])