# -*- coding: utf-8 -*-
"""
This file including functions to generate different types of outliers given
the input dataset for benchmarking
"""
# Author: Yingtong Dou <ytongdou@gmail.com>, Kay Liu <zliu234@uic.edu>
# License: BSD 2 clause
import torch
from torch_geometric.data import Data
from ..utils.utility import check_parameter
[docs]
def gen_structural_outlier(data, m, n, p=0, directed=False, seed=None):
"""Generating structural outliers according to paper :
cite:`ding2019deep`. We randomly select ``m`` nodes from the network
and then make those nodes fully connected, and then all the ``m``
nodes in the clique are regarded as outliers. We iteratively repeat
this process until a number of ``n`` cliques are generated and the
total number of structural outliers is ``m * n``.
Parameters
----------
data : torch_geometric.data.Data
The input data.
m : int
Number nodes in the outlier cliques.
n : int
Number of outlier cliques.
p : int, optional
Probability of edge drop in cliques. Default: ``0``.
directed : bool, optional
Whether the edges added are directed. Default: ``False``.
seed : int, optional
The seed to control the randomness, Default: ``None``.
Returns
-------
data : torch_geometric.data.Data
The structural outlier graph with injected edges.
y_outlier : torch.Tensor
The outlier label tensor where 1 represents outliers and 0
represents normal nodes.
"""
if not isinstance(data, Data):
raise TypeError("data should be torch_geometric.data.Data")
if isinstance(m, int):
check_parameter(m, low=0, high=data.num_nodes, param_name='m')
else:
raise ValueError("m should be int, got %s" % m)
if isinstance(n, int):
check_parameter(n, low=0, high=data.num_nodes, param_name='n')
else:
raise ValueError("n should be int, got %s" % n)
check_parameter(m * n, low=0, high=data.num_nodes, param_name='m*n')
if seed:
torch.manual_seed(seed)
new_edges = []
outlier_idx = torch.randperm(data.num_nodes)[:m * n]
# connect all m nodes in each clique
for i in range(n):
new_edges.append(torch.combinations(outlier_idx[m * i: m * (i + 1)]))
new_edges = torch.cat(new_edges)
# drop edges with probability p
if p != 0:
indices = torch.randperm(len(new_edges))[:int((1-p) * len(new_edges))]
new_edges = new_edges[indices]
y_outlier = torch.zeros(data.x.shape[0], dtype=torch.long)
y_outlier[outlier_idx] = 1
if not directed:
new_edges = torch.cat([new_edges, new_edges.flip(1)], dim=0)
data.edge_index = torch.cat([data.edge_index, new_edges.T], dim=1)
return data, y_outlier
[docs]
def gen_contextual_outlier(data, n, k, seed=None):
r"""Generating contextual outliers according to paper
:cite:`ding2019deep`. We randomly select ``n`` nodes as the
attribute perturbation candidates. For each selected node :math:`i`,
we randomly pick another ``k`` nodes from the data and select the
node :math:`j` whose attributes :math:`x_j` deviate the most from
node :math:`i`'s attribute :math:`x_i` among ``k`` nodes by
maximizing the Euclidean distance :math:`\| x_i − x_j \|`.
Afterwards, we then substitute the attributes :math:`x_i` of node
:math:`i` to :math:`x_j`.
Parameters
----------
data : torch_geometric.data.Data
The input data.
n : int
Number of nodes converting to outliers.
k : int
Number of candidate nodes for each outlier node.
seed : int, optional
The seed to control the randomness, Default: ``None``.
Returns
-------
data : torch_geometric.data.Data
The contextual outlier graph with modified node attributes.
y_outlier : torch.Tensor
The outlier label tensor where 1 represents outliers and 0
represents normal nodes.
"""
if not isinstance(data, Data):
raise TypeError("data should be torch_geometric.data.Data")
if isinstance(n, int):
check_parameter(n, low=0, high=data.num_nodes, param_name='n')
else:
raise ValueError("n should be int, got %s" % n)
if isinstance(k, int):
check_parameter(k, low=0, high=data.num_nodes - n, param_name='k')
else:
raise ValueError("k should be int, got %s" % k)
if seed:
torch.manual_seed(seed)
outlier_idx = torch.randperm(data.num_nodes)[:n]
for i, idx in enumerate(outlier_idx):
candidate_idx = torch.randperm(data.num_nodes)[:k]
euclidean_dist = torch.cdist(data.x[idx].unsqueeze(0), data.x[
candidate_idx])
max_dist_idx = torch.argmax(euclidean_dist, dim=1)
max_dist_node = candidate_idx[max_dist_idx]
data.x[idx] = data.x[max_dist_node]
y_outlier = torch.zeros(data.x.shape[0], dtype=torch.long)
y_outlier[outlier_idx] = 1
return data, y_outlier
# TODO add gen_joint_structural_outliers from GAD-NR