# -*- coding: utf-8 -*-
"""AnomalyDAE: Dual autoencoder for anomaly detection
on attributed networks"""
import warnings
# Author: Xueying Ding <xding2@andrew.cmu.edu>,
# Kay Liu <zliu234@uic.edu>
# License: BSD 2 clause
import torch
import torch.nn.functional as F
from . import DeepDetector
from ..nn import AnomalyDAEBase
[docs]
class AnomalyDAE(DeepDetector):
"""
Dual Autoencoder for Anomaly Detection on Attributed Networks
AnomalyDAE is an anomaly detector that consists of a structure
autoencoder and an attribute autoencoder to learn both node
embedding and attribute embedding jointly in latent space. The
structural autoencoder uses Graph Attention layers. The
reconstruction mean square error of the decoders are defined as
structure anomaly score and attribute anomaly score, respectively,
with two additional penalties on the reconstructed adj matrix and
node attributes (force entries to be nonzero).
See :cite:`fan2020anomalydae` for details.
Parameters
----------
emb_dim : int, optional
Embedding dimension of model. Default: ``64``.
hid_dim : int, optional
Hidden dimension of model. Default: ``64``.
num_layers : int, optional
Total number of layers of AnomalyDAE is fixed to be 4. Changing
of this parameter will not affect the model. Default: ``4``.
dropout : float, optional
Dropout rate. Default: ``0.``.
weight_decay : float, optional
Weight decay (L2 penalty). Default: ``0.``.
act : callable activation function or None, optional
Activation function if not None.
Default: ``torch.nn.functional.relu``.
backbone : torch.nn.Module
The backbone of AnomalyDAE is fixed. Changing of this
parameter will not affect the model. Default: ``None``.
alpha : float, optional
Weight between reconstruction of node feature and structure.
Default: ``0.5``.
theta : float, optional
The additional penalty for nonzero attribute. Default: ``1.``.
eta : float, optional
The additional penalty for nonzero structure. Default: ``1.``.
contamination : float, optional
The amount of contamination of the dataset in (0., 0.5], i.e.,
the proportion of outliers in the dataset. Used when fitting to
define the threshold on the decision function. Default: ``0.1``.
lr : float, optional
Learning rate. Default: ``0.004``.
epoch : int, optional
Maximum number of training epoch. Default: ``100``.
gpu : int
GPU Index, -1 for using CPU. Default: ``-1``.
batch_size : int, optional
Minibatch size, 0 for full batch training. Default: ``0``.
num_neigh : int, optional
Number of neighbors in sampling, -1 for all neighbors.
Default: ``-1``.
verbose : int, optional
Verbosity mode. Range in [0, 3]. Larger value for printing out
more log information. Default: ``0``.
save_emb : bool, optional
Whether to save the embedding. Default: ``False``.
compile_model : bool, optional
Whether to compile the model with ``torch_geometric.compile``.
Default: ``False``.
**kwargs
Other parameters for the backbone model.
Attributes
----------
decision_score_ : torch.Tensor
The outlier scores of the training data. Outliers tend to have
higher scores. This value is available once the detector is
fitted.
threshold_ : float
The threshold is based on ``contamination``. It is the
:math:`N \\times` ``contamination`` most abnormal samples in
``decision_score_``. The threshold is calculated for generating
binary outlier labels.
label_ : torch.Tensor
The binary labels of the training data. 0 stands for inliers
and 1 for outliers. It is generated by applying
``threshold_`` on ``decision_score_``.
emb : torch.Tensor or tuple of torch.Tensor or None
The learned node hidden embeddings of shape
:math:`N \\times` ``hid_dim``. Only available when ``save_emb``
is ``True``. When the detector has not been fitted, ``emb`` is
``None``. When the detector has multiple embeddings,
``emb`` is a tuple of torch.Tensor.
"""
def __init__(self,
emb_dim=64,
hid_dim=64,
num_layers=4,
dropout=0.,
weight_decay=0.,
act=F.relu,
backbone=None,
alpha=0.5,
theta=1.,
eta=1.,
contamination=0.1,
lr=0.004,
epoch=5,
gpu=-1,
batch_size=0,
num_neigh=-1,
verbose=0,
save_emb=False,
compile_model=False,
**kwargs):
if backbone is not None or num_layers != 4:
warnings.warn("Backbone and num_layers are not used in AnomalyDAE")
super(AnomalyDAE, self).__init__(hid_dim=hid_dim,
num_layers=num_layers,
dropout=dropout,
weight_decay=weight_decay,
act=act,
backbone=backbone,
contamination=contamination,
lr=lr,
epoch=epoch,
gpu=gpu,
batch_size=batch_size,
num_neigh=num_neigh,
verbose=verbose,
save_emb=save_emb,
compile_model=compile_model,
**kwargs)
self.emb_dim = emb_dim
self.alpha = alpha
self.theta = theta
self.eta = eta
def process_graph(self, data):
AnomalyDAEBase.process_graph(data)
def init_model(self, **kwargs):
if self.save_emb:
self.emb = torch.zeros(self.num_nodes,
self.hid_dim)
return AnomalyDAEBase(in_dim=self.in_dim,
num_nodes=self.num_nodes,
emb_dim=self.emb_dim,
hid_dim=self.hid_dim,
dropout=self.dropout,
act=self.act,
**kwargs).to(self.device)
def forward_model(self, data):
batch_size = data.batch_size
node_idx = data.n_id
x = data.x.to(self.device)
s = data.s.to(self.device)
edge_index = data.edge_index.to(self.device)
x_, s_ = self.model(x, edge_index, batch_size)
# positive weight conversion
weight = 1 - self.alpha
pos_weight_a = self.eta / (1 + self.eta)
pos_weight_s = self.theta / (1 + self.theta)
score = self.model.loss_func(x[:batch_size],
x_[:batch_size],
s[:batch_size, node_idx],
s_[:batch_size],
weight,
pos_weight_a,
pos_weight_s)
loss = torch.mean(score)
return loss, score.detach().cpu()