faknow.model.social_context

faknow.model.social_context.base_gnn

class faknow.model.social_context.base_gnn.GAT(feature_size: int, hidden_size=128)[source]

Bases: _BaseGNN

Graph Attention Networks, ICLR 2018 paper: https://openreview.net/forum?id=rJXMpikCZ code: https://github.com/safe-graph/GNN-FakeNews

__init__(feature_size: int, hidden_size=128)[source]

Parameters:

feature_size (int) – dimension of input node feature
hidden_size (int) – Default=128

training: bool

class faknow.model.social_context.base_gnn.GCN(feature_size: int, hidden_size=128)[source]

Bases: _BaseGNN

Semi-Supervised Classification with Graph Convolutional Networks, ICLR 2017 paper: https://openreview.net/forum?id=SJU4ayYgl code: https://github.com/safe-graph/GNN-FakeNews

__init__(feature_size: int, hidden_size=128)[source]

Parameters:

feature_size (int) – dimension of input node feature
hidden_size (int) – Default=128

training: bool

class faknow.model.social_context.base_gnn.SAGE(feature_size: int, hidden_size=128)[source]

Bases: _BaseGNN

Inductive Representation Learning on Large Graphs, NeurIPS 2017 paper: https://dl.acm.org/doi/10.5555/3294771.3294869 code: https://github.com/safe-graph/GNN-FakeNews

__init__(feature_size: int, hidden_size=128)[source]

Parameters:

feature_size (int) – dimension of input node feature
hidden_size (int) – Default=128

training: bool

faknow.model.social_context.bigcn

class faknow.model.social_context.bigcn.BiGCN(feature_size: int, hidden_size=128, out_size=128)[source]

Bases: AbstractModel

Rumor Detection on Social Media with Bi-Directional Graph Convolutional Networks, AAAI 2020 paper: https://ojs.aaai.org/index.php/AAAI/article/view/5393 code: https://github.com/safe-graph/GNN-FakeNews

__init__(feature_size: int, hidden_size=128, out_size=128)[source]

Parameters:

feature_size (int) – dimension of input node feature
hidden_size (int) – dimension of hidden layer. Default=128
out_size (int) – dimension of output layer. Default=128

calculate_loss(data: Batch) → Tensor[source]

calculate loss via CrossEntropyLoss

Parameters:: data (Batch) – batch data
Returns:: loss
Return type:: torch.Tensor

forward(x: Tensor, edge_index: Tensor, bu_edge_index: Tensor, batch: Tensor, root_index: Tensor)[source]

Parameters:

x (Tensor) – node feature, shape=(num_nodes, feature_size), shape=(num_nodes, feature_size)
edge_index (Tensor) – top-down edge index, shape=(2, num_edges)
bu_edge_index (Tensor) – bottom-up edge index, shape=(2, num_edges)
batch (Tensor) – index of graph each node belongs to, shape=(num_nodes,)
root_index (Tensor) – index of root node for each graph, shape=(num_graphs,)

Returns:

prediction of being fake, shape=(batch_size, 2)

Return type:

Tensor

predict(data_without_label) → Tensor[source]

predict the probability of being fake news

Parameters:: data_without_label (Batch) – batch data
Returns:: softmax probability, shape=(num_graphs, 2)
Return type:: Tensor

training: bool

faknow.model.social_context.dudef

class faknow.model.social_context.dudef.DUDEF(input_size: int, emotion_len: int, hidden_size: int, embedding_matrix: Tensor)[source]

Bases: AbstractModel

DUDEF: Mining Dual Emotion for Fake News Detection, WWW 2021 paper: https://dl.acm.org/doi/10.1145/3442381.3450004 code: https://github.com/RMSnow/WWW2021

__init__(input_size: int, emotion_len: int, hidden_size: int, embedding_matrix: Tensor)[source]

Parameters:

input_size (int) – dimension of input node feature
emotion_len (int) – dimension of dual_emotion
hidden_size (int) – dimension of hidden layer
embedding_matrix (Tensor) – word embedding matrix

calculate_loss(data: Dict[str, Tensor]) → Tensor[source]

calculate loss via CrossEntropyLoss

Parameters:: data (Dict[str, Tensor]) – batch data, shape=(2,batch_size,feature_size)
Returns:: loss
Return type:: Tensor

forward(semantics: Tensor, emotions: Tensor) → Tensor[source]

Parameters:

semantics (Tensor) – node semantics features, shape=(batch_size, feature_size)
emotions (Tensor) – node emotions features, shape=(batch_size, feature_size)

Returns:

prediction of being fake, shape=(batch_size, 2)

Return type:

Tensor

predict(data: Dict[str, Tensor]) → Tensor[source]

predict the probability of being fake news

Parameters:: data (Batch) – batch data, shape=(2,batch_size,feature_size)
Returns:: softmax probability, shape=(batch_size, 2)
Return type:: Tensor

training: bool

faknow.model.social_context.ebgcn

class faknow.model.social_context.ebgcn.EBGCN(input_size=5000, hidden_size=64, output_size=64, edge_num=2, dropout=0.5, num_class=4, edge_loss_weight=0.2)[source]

Bases: AbstractModel

Towards Propagation Uncertainty: Edge-enhanced Bayesian Graph Convolutional Networks for Rumor Detection, ACL 2021 paper: https://aclanthology.org/2021.acl-long.297/ code: https://github.com/weilingwei96/EBGCN

__init__(input_size=5000, hidden_size=64, output_size=64, edge_num=2, dropout=0.5, num_class=4, edge_loss_weight=0.2)[source]

Parameters:

input_size (int) – the feature size of input. default=5000.
hidden_size (int) – the feature size of hidden embedding. default=64.
output_size (int) – the feature size of output embedding. default=64.
edge_num (int) – the num of edge type. default=2.
dropout (float) – dropout rate. default=0.5.
num_class (int) – the num of output type. default=4
edge_loss_weight (float) – the weight of edge loss. default=0.2.

calculate_loss(data: Batch) → Dict[str, float | Any][source]

calculate loss for EBGCN

Parameters:: data (Batch) – pyg batch data
Returns:: loss dict with key ‘total_loss’, ‘pred_loss’, ‘edge_loss’
Return type:: dict

forward(node_features: Tensor, td_edge_index: Tensor, bu_edge_index: Tensor, root_index: Tensor, batch: Tensor)[source]

Parameters:

node_features (Tensor) – feature of node in claims.
td_edge_index (Tensor) – directed adjacent martix in COO format from top to bottom.
bu_edge_index (Tensor) – directed adjacent martix in COO format from bottom to top.
root_index (Tensor) – index of root news in claims.
batch (Tensor) – vector mapping each node to its respective graph.

Returns:

predict output. td_edge_loss(Tensor): edge loss of graph with direction from top to bottom. bu_edge_loss(Tensor): edge loss of graph with direction from bottom to top.

Return type:

output(Tensor)

predict(data_without_label: Batch) → Tensor[source]

predict the probability of being fake news

Parameters:: data_without_label (Batch) – pyg batch data
Returns:: softmax probability, shape=(batch_size, num_classes)
Return type:: Tensor

training: bool

faknow.model.social_context.eddfn

class faknow.model.social_context.eddfn.EDDFN(input_size: int, domain_size: int, lambda1=1.0, lambda2=10.0, lambda3=5.0, hidden_size=512)[source]

Bases: AbstractModel

Embracing Domain Differences in Fake News Cross-domain Fake News Detection using Multi-modal Data, AAAI 2021 paper: https://ojs.aaai.org/index.php/AAAI/article/view/16134 code: https://github.com/amilasilva92/cross-domain-fake-news-detection-aaai2021

__init__(input_size: int, domain_size: int, lambda1=1.0, lambda2=10.0, lambda3=5.0, hidden_size=512)[source]

Parameters:

input_size (int) – dimension of input representation
domain_size (int) – dimension of domain vector
lambda1 (float) – L_{recon} loss weight. Default=1.0
lambda2 (float) – L_{specific} loss weight. Default=10.0
lambda3 (float) – L_{shared} loss weight. Default=5.0
hidden_size (int) – size of hidden layer. Default=512

calculate_loss(data: Tuple[Tensor]) → Dict[str, Tensor][source]

calculate total loss, including classification loss(BCELoss), reconstruction loss(MSELoss), specific domain loss(MSELoss) and shared domain loss(MSELoss)

Parameters:: data (Tuple[Tensor]) – batch data tuple, including input, domain and label
Returns:: loss dict, key: total_loss, class_loss, decoder_loss, specific_domain_loss, shared_domain_loss
Return type:: Dict[str, Tensor]

forward(input_representation: Tensor) → Tuple[Tensor, Tensor, Tensor, Tensor][source]

Parameters:: input_representation (Tensor) – input representation consisting of text and propagation network
Returns:: class_out (Tensor): prediction of being fake news, shape=(batch_size, 1) decoder_out (Tensor): prediction of input, shape=(batch_size, input_size) specific_domain (Tensor): specific domain output, shape=(batch_size, domain_size) shared_domain (Tensor): shared domain output, shape=(batch_size, domain_size)
Return type:: tuple

predict(data_without_label: Tuple[Tensor])[source]

predict the probability of being fake news

Parameters:: data_without_label (Tuple[Tensor]) – batch data tuple, including input, domain
Returns:: one hot probability, shape=(batch_size, 2)
Return type:: Tensor

training: bool

faknow.model.social_context.fang

class faknow.model.social_context.fang.FANG(fang_data: FangDataset, input_size=100, embedding_size=16, num_stance=4, num_stance_hidden=4, timestamp_size=2, num_classes=2, dropout=0.0, device='cpu')[source]

Bases: AbstractModel

FANG: Leveraging Social Context for Fake News Detection Using Graph Representation, CIKM 2020 paper: https://dl.acm.org/doi/10.1145/3340531.3412046 code: https://github.com/nguyenvanhoang7398/FANG

__init__(fang_data: FangDataset, input_size=100, embedding_size=16, num_stance=4, num_stance_hidden=4, timestamp_size=2, num_classes=2, dropout=0.0, device='cpu')[source]

Parameters:

fang_data (FangDataset) – global graph information.
input_size (int) – raw features’ embedding size. default=100.
embedding_size (int) – embedding size of fang and the output size of graphsage. default=16.
num_stance (int) – the total num of stance. default=4.
num_stance_hidden (int) – the feature size of stance(usually num_stance_hidden * num_stance = embedding_size). default=4.
timestamp_size (int) – the feature size of timestamp. default=2.
num_classes (int) – label num. default=2.
dropout (float) – dropout rate. default=0.1.

calculate_loss(data: Tensor) → dict[source]

calculate total loss

Parameters:: data (torch.Tensor) – nodes’ idx list
Returns:: news_loss + stance_loss + unsup_loss
Return type:: Torch.Tensor

extract_most_attended_users(news_attn: Tensor, engage_users: list)[source]

using attention to search the most attended user to compute unsupvised loss.

Parameters:

news_attn (torch.Tensor) – users’ attention for news.
engage_users (list) – all engaged users.

fetch_news_user_stance(news_nodes: list, stance_lists: dict)[source]

fetch data required for stance prediction.

Parameters:

news_nodes (list) – list of news node.
stance_lists (dict) – stance information dict for every node.

forward(data_batch: list, test=False)[source]

Parameters:

data_batch (list) – node’s idx list.
test (bool) – test or train. default=False

get_neigh_weights(node: int, node_only=False)[source]

get node’s neighbor.

Parameters:

node (int) – node idx.
node_only (bool) – whether to get only node or both node and its edge information. default=False.

get_proximity_samples(nodes: list, num_pos: int, num_neg: int, neg_walk_len: int)[source]

positive samping and negative sampling.

Parameters:

nodes (list) – node idx list.
num_pos (int) – positive sampling num.
num_neg (int) – negative sampling num.
neg_walk_len (int) – walk step.

infer_embedding(node, sample_size)[source]

predict(data_without_label: dict) → Tensor[source]

predict the probability of being fake news

Parameters:: data_without_label (dict) – batch data
Returns:: predict probability, shape=(batch_size, 2)
Return type:: torch.Tensor

preprocess_news_classification_data(nodes_batch: list, embed_size: int, test=False)[source]

Parameters:

nodes_batch (list) – the idx list of node to be processed.
embed_size (int) – the size of embedding features.
test (bool) – if test or trai

training: bool

faknow.model.social_context.gcnfn

class faknow.model.social_context.gcnfn.GCNFN(feature_size: int, hidden_size=128, dropout_ratio=0.5)[source]

Bases: _BaseGCNFN

Fake news detection on social media using geometric deep learning, arXiv 2019 paper: https://arxiv.org/abs/1902.06673 code: https://github.com/safe-graph/GNN-FakeNews

__init__(feature_size: int, hidden_size=128, dropout_ratio=0.5)[source]

Parameters:

feature_size (int) – dimension of input node feature
hidden_size (int) – dimension of hidden layer. Default=128
dropout_ratio (float) – dropout ratio. Default=0.5

training: bool

faknow.model.social_context.gnncl

class faknow.model.social_context.gnncl.GNNCL(feature_size: int, max_nodes: int)[source]

Bases: AbstractModel

Graph neural networks with continual learning for fake news detection from social media, arXiv 2020 paper: https://arxiv.org/abs/2007.03316 code: https://github.com/safe-graph/GNN-FakeNews

__init__(feature_size: int, max_nodes: int)[source]

Parameters:

feature_size (int) – dimension of input node feature
max_nodes (int) – maximum number of nodes in a graph

calculate_loss(data: Batch) → Tensor[source]

calculate loss via CrossEntropyLoss

Parameters:: data (Batch) – batch data
Returns:: loss
Return type:: torch.Tensor

forward(x: Tensor, adj: Tensor, mask: Tensor | None = None)[source]

Parameters:

x (Tensor) – node features, shape=(batch_size, max_nodes, feature_size)
adj (Tensor) – adjacency matrix, shape=(batch_size, max_nodes, max_nodes)
mask (Tensor) – mask for adjacency matrix, shape=(batch_size, max_nodes). Default=None.

Returns:

output (Tensor): prediction of being fake, shape (batch_size, 2) l (Tensor): loss of first pooling layer e (Tensor): entropy of first pooling layer

Return type:

tuple

predict(data_without_label) → Tensor[source]

predict the probability of being fake news

Parameters:: data_without_label (Batch) – batch data
Returns:: softmax probability, shape=(batch_size, 2)
Return type:: Tensor

training: bool

faknow.model.social_context.trustrd

class faknow.model.social_context.trustrd.TrustRD(pre_trained_model: TrustRDPreTrainModel, in_feature=192, hid_feature=64, num_classes=4, sigma_m=0.1, eta=0.4, zeta=0.02)[source]

Bases: AbstractModel

Towards Trustworthy Rumor Detection with Interpretable Graph Structural Learning, CIKM 2023 paper: https://dl.acm.org/doi/10.1145/3583780.3615228 code: https://github.com/Anonymous4ScienceAuthor/TrustRD

__init__(pre_trained_model: TrustRDPreTrainModel, in_feature=192, hid_feature=64, num_classes=4, sigma_m=0.1, eta=0.4, zeta=0.02)[source]

Parameters:

pre_trained_model (TrustRDPreTrainModel) – the model of trained TrustRDPreTrainModel.
in_feature (int) – the feature size of input. default=192.
hid_feature (int) – the feature size of hidden embedding. default=64.
num_classes (int) – the num of class. default=4.
sigma_m (float) – data perturbation Standard Deviation. default=0.1
eta (float) – data perturbation weight. default=0.4.
zeta (float) – parameter perturbation weight. default=0.02

calculate_loss(data: Batch) → Dict[str, float][source]

loss = pred_loss + 0.2 * loss_para + 0.2 * loss_data, where pred_loss is the cross entropy loss, loss_para is the parameter perturbation loss, and loss_data is the data perturbation loss.

Parameters:: data (Batch) – batch data
Returns:: loss dict with key ‘total_loss’, ‘pred_loss’, ‘para_loss’, ‘data_loss’
Return type:: dict

forward(embed: Tensor, data: Data)[source]

Parameters:

embed (Tensor) – embedding features processed by TrustRDPreTrainModel.
data (Data) – batch data.

Returns:

[predict_result, last_prob, kl_div]

Return type:

Tensor

predict(data_without_label: Batch) → Tensor[source]

predict the probability of being fake news

Parameters:: data_without_label (Batch) – batch data
Returns:: predication probability, shape=(num_graphs, 2)
Return type:: Tensor

training: bool

weights_init(m)[source]

class faknow.model.social_context.trustrd.TrustRDPreTrainModel(hidden_dim=64, num_gcn_layers=3, num_classes=4)[source]

Bases: AbstractModel

Self-Supervised Pretraining Model for TrustRD

__init__(hidden_dim=64, num_gcn_layers=3, num_classes=4)[source]

Parameters:

hidden_dim (int) – the feature size of hidden embedding. default=64.
num_gcn_layers (int) – the gcn pre_trained_model layer num. default=3.
num_classes (int) – the num of class. default=4.

calculate_loss(data: Batch) → Tensor[source]

calculate ib_loss

Parameters:: data (Batch) – batch data
Returns:: ib_loss
Return type:: torch.Tensor

forward(data: Batch) → Tuple[Tensor, Tensor, Tensor][source]

Parameters:

data (Batch) – pyg data batch

Returns:

positive augment data, negative augment data: and average representation

Return type:

Tuple[Tensor, Tensor, Tensor]

generate_drop_edge(x: Tensor, edge_index: Tensor)[source]: Generate dropped edge to augment data.

init_emb()[source]

training: bool

faknow.model.social_context.trustrd.calculate_ib_loss(g_enc_pos: Tensor, g_enc_neg: Tensor, g_enc: Tensor) → Tensor[source]

Compute contrastive loss :param g_enc_pos: positive augment data. :type g_enc_pos: Tensor :param g_enc_neg: negative augment data. :type g_enc_neg: Tensor :param g_enc: average representation. g_enc = (g_enc_pos + g_enc_neg) / 2. :type g_enc: Tensor

Returns:: ib_loss
Return type:: Tensor

faknow.model.social_context.upfd

class faknow.model.social_context.upfd.UPFDGAT(feature_size: int, hidden_size=128)[source]

Bases: _BaseGNN

User Preference-aware Fake News Detection, SIGIR 2021 paper: https://dl.acm.org/doi/abs/10.1145/3404835.3462990 code: https://github.com/safe-graph/GNN-FakeNews

UPFD model with GAT layer

__init__(feature_size: int, hidden_size=128)[source]

Parameters:

feature_size (int) – dimension of input node feature
hidden_size (int) – dimension of hidden layer. Default=128

training: bool

class faknow.model.social_context.upfd.UPFDGCN(feature_size: int, hidden_size=128)[source]

Bases: _BaseGNN

User Preference-aware Fake News Detection, SIGIR 2021 paper: https://dl.acm.org/doi/abs/10.1145/3404835.3462990 code: https://github.com/safe-graph/GNN-FakeNews

UPFD model with GCN layer

__init__(feature_size: int, hidden_size=128)[source]

Parameters:

feature_size (int) – dimension of input node feature
hidden_size (int) – dimension of hidden layer. Default=128

training: bool

class faknow.model.social_context.upfd.UPFDGCNFN(feature_size: int, hidden_size=128, dropout_ratio=0.5)[source]

Bases: _BaseGCNFN

User Preference-aware Fake News Detection, SIGIR 2021 paper: https://dl.acm.org/doi/abs/10.1145/3404835.3462990 code: https://github.com/safe-graph/GNN-FakeNews

UPFD model with GCNFN layer

__init__(feature_size: int, hidden_size=128, dropout_ratio=0.5)[source]

Parameters:

feature_size (int) – dimension of input node feature
hidden_size (int) – dimension of hidden layer. Default=128
dropout_ratio (float) – dropout ratio. Default=0.5

training: bool

class faknow.model.social_context.upfd.UPFDSAGE(feature_size: int, hidden_size=128)[source]

Bases: _BaseGNN

User Preference-aware Fake News Detection, SIGIR 2021 paper: https://dl.acm.org/doi/abs/10.1145/3404835.3462990 code: https://github.com/safe-graph/GNN-FakeNews

UPFD model with SAGE layer

__init__(feature_size: int, hidden_size=128)[source]

Parameters:

feature_size (int) – dimension of input node feature
hidden_size (int) – dimension of hidden layer. Default=128

training: bool