「论文学习」DeepInf: Social Influence Prediction with Deep Learning

2020-07-18

Basic Info

Link

DeepInf: Social Influence Prediction with Deep Learning

Author

Jiezhong Qiu
Jian Tang
Hao Ma
Yuxiao Dong
Kuansan Wang
Jie Tang

Method(s)

Architecture

(a): Raw input which consists of a mini-batch of B instances; Each instance is a sub-network comprised n users who are sampled from the whole network. (b): An embedding layer which maps each user to a D-dimensional representation. (c): An Instance Normalization Layer which normalizes user’s embedding. (d): The formal input layer which concatenates together network embedding and other features. (e) A GCN and GAT Layer. (f) and (g) Compare model output and ground truth, we get the negative log likelihood loss. In this example, ego user v was finally activated (marked as black)

Reproduce

代码仓库

数据导入

对于npy格式的数据

使用numpy的load导入

1
2
3

import numpy as np

data_np = np.load(data_path)

对于emb_64的数据

这个我不了解格式，参考了baseline中的代码

def load_w2v_feature(file, max_idx=0):
    with open(file, "rb") as f:
        nu = 0
        for line in f:
            content = line.strip().split()
            nu += 1
            if nu == 1:
                n, d = int(content[0]), int(content[1])
                feature = [[0.] * d for i in range(max(n, max_idx + 1))]
                continue
            index = int(content[0])
            while len(feature) <= index:
                feature.append([0.] * d)
            for i, x in enumerate(content[1:]):
                feature[index][i] = float(x)
    for item in feature:
        assert len(item) == d
    return np.array(feature, dtype=np.float32)

邻接矩阵

针对邻接矩阵，自身到自己的距离可以设为1，因此可以用单位矩阵加上数据中的adjacency matrix

adjacency_matrix_filename = 'adjacency_matrix.npy'
adjacency_matrix_filepath = os.path.join(dir_path, adjacency_matrix_filename)
self.graphs = adjacency_matrix_np.astype(adjacency_matrix_filepath)
        
identity = np.identity(self.graphs.shape[1])
self.graphs += identity