CasSeqGCN: Combining network structure and temporal sequence to predict information cascades

CasTemporalGCN: early cascade growth prediction with considering temporal features based on graph convolutional networks

“…The main features used include content features, user features, temporal features, and structural features. Traditional methods mostly use heuristics to extract features, and these methods are generally difficult to transfer across different datasets, and their feature extraction capabilities are inferior to deep learning methods [3,8,21] .…”

“…Li et al [4] first applied deep learning to the cascade growth prediction task and proposed the DeepCas model; Li et al [22] proposed a DCGT model based on DeepCas, which incorporates cascading content features into node features through a gating mechanism; Cao et al 's [5] DeepHawkes model combines the Hawkes point process model and deep learning, using GRU to automatically learn the user influence vector in the Hawkes process, with strong interpretability; Kafato et al [23] used the CAS2VEC model, which discretizes temporal information, to predict whether the cascade will propagate widely using convolutional neural networks. Su et al [24] improved the DeepWalk method by proposing a new random walking algorithm on heterogeneous graphs for the cascade growth prediction problem, HWalk, and applied it to the heterogeneous networks formed by cascades and users; The NPP model proposed by Chen et al [25] focuses on the extraction and fusion of multifeature information; Wang et al successively proposed CasSeqGCN [3] and CCasGNN [8] . CCasGNN encodes the location information by positive cosine vectors and uses GAT and GCN to collaboratively predict length.…”

“…In order to evaluate the performance of our method thoroughly, we carry out experiment on the synthetic dataset and real datasets of Sina Weibo [28,29] and ACM citation [30] . The synthetic datasets were generated using the Barabasi-Albert model to construct social networks with parameters specified by Wang et al [3,9] , and the Independent Cascade model to simulate information propagation and generate cascades of information using the method described by Cao et al [9] . Figure 2 shows the distribution of concatenation lengths-the number of times each message is forwarded-following a power-law distribution.…”

“…In this study, we evaluated several state-of-the-art methods for cascade prediction, including the Feature-based method, DeepCas [4] , DeepHawkes [5] , CoupledGNN [9] , CasSeqGCN [3] , and CCasGNN [8] . The Feature-based method uses the same features as input and uses a three-layer fully connected network for prediction, with parameters chosen from {32,64,128}.…”

“…There are a large number of studies on cascade growth prediction, and researchers usually use Recurrent Neural Network (RNN) and variants of RNN [3][4][5][6][7] to extract temporal features of the cascade, extract structural features of the cascade via Graph Neural Network (GNN) [3,[8][9][10][11][12] , and apply other techniques, such as attention mechanisms [13][14][15] , variational inference [16,17] , and reinforcement learning [17] , to jointly predict the cascade growth.…”