Source-linked AI summary
DeepCas: an End-to-end Predictor of Information Cascades
Cheng Li, Jiaqi Ma, Xiaoxiao Guo, Qiaozhu Mei
TL;DR
Cascade prediction has shown promise but depends on difficult-to-generalize hand-crafted features and must represent cascade graphs rather than isolated nodes. DeepCas learns cascade-graph representations end to end from sampled paths, GRU encoding, and attention, and outperforms feature-based, node-embedding, and graph-embedding alternatives. The study also notes limits in interpreting the learned representations and in observing complete cascade histories.
Problem
Existing cascade predictors rely on hand-crafted features, while representing a cascade graph as a whole remains challenging.
Method
DeepCas samples cascade paths and processes them with a GRU-based neural network and attention mechanism to predict future cascade size.
Results
DeepCas outperforms feature-based machine-learning methods and alternative node-embedding and graph-embedding methods.
Takeaways & Limitations
The reported benefit of deep learning comes from its end-to-end procedure, which likely learns high-level features beyond strong hand-crafted representations.
Takeaways & Limitations
The learned representation was not qualitatively assessed for what information it captures, and observations may lack complete cascade histories.
Abstract
from arXiv · showhide
Information cascades, effectively facilitated by most social network platforms, are recognized as a major factor in almost every social success and disaster in these networks. Can cascades be predicted? While many believe that they are inherently unpredictable, recent work has shown that some key properties of information cascades, such as size, growth, and shape, can be predicted by a machine learning algorithm that combines many features. These predictors all depend on a bag of hand-crafting features to represent the cascade network and the global network structure. Such features, always carefully and sometimes mysteriously designed, are not easy to extend or to generalize to a different platform or domain. Inspired by the recent successes of deep learning in multiple data mining tasks, we investigate whether an end-to-end deep learning approach could effectively predict the future size of cascades. Such a method automatically learns the representation of individual cascade graphs in the context of the global network structure, without hand-crafted features and heuristics. We find that node embeddings fall short of predictive power, and it is critical to learn the representation of a cascade graph as a whole. We present algorithms that learn the representation of cascade graphs in an end-to-end manner, which significantly improve the performance of cascade prediction over strong baselines that include feature based methods, node embedding methods, and graph kernel methods. Our results also provide interesting implications for cascade prediction in general.
1. INTRODUCTION
Information cascades can be predicted to some extent, but existing predictors depend heavily on hand-crafted features whose design is difficult to systematize. DeepCas addresses this gap with an end-to-end model that learns cascade-graph representations for future-size prediction.
- Information cascades affect viral marketing, innovation diffusion, crowdsourcing, rumors, cyber violence, and persuasion campaigns.
- Cascade properties such as size, growth, and shape have been predicted across Tweets, photos, videos, and academic papers using machine-learning methods.
- Existing predictors rely on carefully designed and sometimes ad hoc features, including centrality, communities, tie strength, structural holes, and platform-specific signals.
- Feature-based performance depends heavily on representation choices, yet no common principle guides feature design, measurement, or selection.
- DeepCas uses end-to-end deep learning to automatically learn cascade-graph representations that predict future cascade size without arbitrary feature design.
- DeepCas represents cascade graphs as sampled paths and significantly improves prediction accuracy over feature-based, node-embedding, and graph-kernel baselines.
2. RELATED WORK
Prior cascade-prediction methods use generative assumptions or hand-crafted features, while graph-representation methods face difficulties incorporating both structure and node identity. DeepCas instead learns cascade representations end to end for future-size prediction.
- Information cascades describe chain reactions through which opinions, advertisements, or propaganda spread among influenced neighbors.
- Cascade prediction is formulated as classification of popularity or regression of numerical properties such as cascade size.
- Generative cascade models use temporal or structural assumptions that can oversimplify real diffusion processes.
- Hand-crafted feature methods combine extracted network factors with supervised models, but performance depends heavily on feature quality and lacks systematic design principles.
- DeepCas directly learns cascade representations without arbitrary feature design and focuses on cascade structure, including node identities.
- Node embeddings represent individual nodes, whereas graph kernels capture structural similarity but struggle to incorporate node identity information.
- The paper introduces an end-to-end architecture that learns cascade-graph representations to optimize future-size prediction accuracy.
3. METHOD
DeepCas formulates cascade prediction from partially observed network snapshots and learns graph representations end to end from sampled random-walk paths. Its pipeline uses graph-specific sampling, recurrent sequence encoding, and attention to predict future cascade-size increments without relying on fixed path parameters.
- Problem definition: The task predicts a cascade’s size increment after a horizon Δt from a social-network snapshot and an observed cascade graph.Earlier prediction times and longer horizons are valuable but harder; the global network is treated as static.
- Cascade graph representation: DeepCas samples paths rather than individual nodes to preserve local and global structural information in cascade-graph representations.The paths are generated through random walks over the cascade graph.
- Cascade graph representation: Random-walk transitions use scoring functions based on cascade out-degree, global degree, or edge weight, with jumping and termination probabilities controlling path behavior.The transition probabilities determine neighbor selection, while pj corresponds to sequence length and po to the expected number of sampled sequences.
- Sampling sequences from a graph: DeepCas learns graph-specific path lengths and counts instead of fixing the random-walk parameters or using predefined T and K.It samples sufficiently long and numerous sequences, then learns the actual length and count required for each cascade graph.
- Neural network models: Attention assembles encoded sequences into a whole-graph representation that is optimized for predicting future cascade growth.The model learns how many sequences and how much of each sequence to use for each graph.
- Neural network models: A shared embedding matrix converts one-hot node identities into vectors, and a bidirectional GRU encodes information flow in both sequence directions.The forward and backward hidden vectors are concatenated into each node’s sequence representation.
4. EXPERIMENT SETUP
The experiments evaluate DeepCas on Twitter and scientific-paper cascades, using network constructions tailored to each domain and regression of future cascade growth. DeepCas is compared with feature-based, node-embedding, graph-kernel, and alternative deep-learning baselines under specified preprocessing and training settings.
- 4.1 Data Sets: The evaluation uses Twitter retweet cascades and scientific-paper citation cascades from the TWITTER and AMINER data sets.Twitter data are split temporally into training, validation, and testing periods; the paper network is constructed from citations between 1992 and 2002.
- 4.1 Data Sets: The Twitter global network connects users when retweets or mentions provide evidence of active interaction, because follower relations and retweet paths are unavailable.Edges are drawn from A to B if B retweeted A’s message or A mentioned B.
- 4.2 Prediction Task: 50% of zero-growth graphs are downsampled, and the target is transformed as yi = log2(∆si + 1) for regression evaluation.Prediction accuracy is measured with mean squared error (MSE).
- 4.3 Baseline Methods: DeepCas is compared with feature-based methods, node-embedding methods, alternative deep-learning graph representations, and graph-kernel baselines.Feature baselines include structural, identity, community, and substructure features; Node2vec averages node embeddings before MLP prediction.
- 4.3 Baseline Methods: Node2vec combines cascade-graph and global-network walks into node embeddings, whose cascade-level average is passed through an MLP.Embedded-IC provides another embedding-based comparison, while PSCN applies CNNs to locally connected graph regions.
- 4.4 DeepCas Variants: DeepCas variants use random-walk transition probabilities based on edge weights, cascade-graph degree, or global-network degree, while GRU-bag tests whether sequential information matters.The main configuration samples K = 200 paths of length T = 10; attention determines when and where sequences stop contributing.
5. EXPERIMENT RESULTS
DeepCas outperforms the eight baseline methods and reveals that end-to-end path-based graph representations capture structural and global network properties. Its advantages are strongest for larger, denser cascades, although the learned representations remain difficult to interpret.
- 5.1 Overall performance: DeepCas models significantly outperform all eight baselines in MSE across datasets.Table 2 reports errors on log-transformed outcomes; differences would appear larger after translating them to raw cascade sizes.
- 5.1 Overall performance: End-to-end learning matters more than model depth when strong hand-crafted features are already available.Features-deep does not always outperform Features-linear, highlighting the value of learning representations directly from data.
- 5.1 Overall performance: Averaging node embeddings is less informative than representing cascade graphs as sampled paths, while node identities create sparsity problems for graph kernels.Node2Vec and Embedded-IC perform poorly, and WL-id degenerates because graph-kernel similarity based on node IDs is sparse.
- 5.1 Overall performance: Path sampling, adaptive walk control, and random jumps beyond the root each contribute to DeepCas performance.Removing these components in GRU-bag, GRU-fixed, and GRU-root causes degradation; random jumps also help handle missing data.
- 5.1 Overall performance: Larger observation times improve prediction, whereas longer prediction horizons make prediction harder across methods.More elapsed cascade information helps, while forecasting farther into the future is more difficult.
- 5.2 Interpreting the Representations: The qualitative meaning of DeepCas representations remains unassessed, leaving interpretability as an open concern.The authors describe deep-learning models as black boxes and do not establish exactly what the learned representation captures.
6. DISCUSSION AND CONCLUSION
DeepCas learns cascade-graph representations end to end and outperforms feature-based, node-embedding, and graph-embedding alternatives. The discussion highlights when hand-crafted features remain strong, how random-walk strategies vary across scenarios, and the scope for adding other information.
- DeepCas represents cascade graphs with random-walk paths, a GRU, and attention to predict future cascade size.
- Hand-crafted features remain reasonably strong and interpretable, while DeepCas’s benefit is attributed to end-to-end learning of higher-level network representations.
- Different random-walk strategies perform differently across scenarios, and all outperform bags of node embeddings.
- The study uses only network structure and node identities, leaving content and time series as information that could be incorporated for domain-specific accuracy.