Research map / Graph learning / Graph self-supervised learning
MSR: A Multifaceted Self-Retrieval Framework for Microscopic Cascade Prediction
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MSR, a microscopic cascade prediction framework that adds a behavioral preference graph to the social graph, uses a multi-channel GRAU encoder, and replaces classification with a cosine-similarity self-retrieval decoder.
Problem
Microscopic cascade prediction (predicting the next user to be infected in an information cascade) is usually cast as N-class classification that uses only infected users' features and mainly models social relationships, ignoring uninfected users and behavioral preferences.
Approach
Builds a preference graph from users' historical participation alongside the social graph, encodes infected users with a multi-channel GRAU, and recasts prediction as retrieval: the next user is the one whose composite feature has highest cosine similarity to the cascade hidden state, leveraging Social Homophily over uninfected users too.
Key results
- MSR reaches 65.08 Hits@100 on Twitter vs 60.41 for GRASS, and 75.39 Hits@100 on Memetracker vs 68.34 for GRASS.
- Averaged over datasets, MSR improves Hits@κ by 6.0% and MAP@κ by 4.4% relative to baselines (against GRASS: 4.46%, 7.17%, 6.38% Hits gains).
- Ablation shows removing self-retrieval drops Memetracker Hits@100 by 4.21%, and single-graph variants underperform the full model.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MSR instead of GRASS | Microscopic cascade prediction · Twitter | Hits@100: 60.41 -> 65.08 | Helped |
| Use MSR instead of GRASS | Microscopic cascade prediction · Twitter | MAP@100: 22.57 -> 24.00 | Helped |
| Use MSR instead of GRASS | Microscopic cascade prediction · Memetracker | Hits@100: 68.34 -> 75.39 | Helped |
| Use MSR instead of GRASS | Microscopic cascade prediction · Memetracker | MAP@100: 27.97 -> 34.82 | Helped |
| Use MSR instead of GRASS | Microscopic cascade prediction · Douban | Hits@100: 44.92 -> 50.99 | Helped |
| Use MSR instead of GRASS | Microscopic cascade prediction · Douban | MAP@100: 15.16 -> 20.12 | Helped |
| Add self-retrieval decoding (ablation) | MSR · Twitter | Hits@100: 63.07 -> 65.08 | Helped |
| Add self-retrieval decoding (ablation) | MSR · Twitter | MAP@100: 23.37 -> 24.00 | Helped |
| Add self-retrieval decoding (ablation) | MSR · Memetracker | Hits@100: 71.18 -> 75.39 | Helped |
| Add self-retrieval decoding (ablation) | MSR · Memetracker | MAP@100: 31.17 -> 34.82 | Helped |
| Add multi-channel encoding rather than using the union graph (ablation) | MSR · Twitter | MAP@100: 22.61 -> 24.00 | Helped |
| Add multi-channel encoding rather than using the union graph (ablation) | MSR · Memetracker | MAP@100: 33.84 -> 34.82 | Helped |
Surprising findings
- On Twitter, the self-retrieval decoder alone improves Hits@100 but only matches GRASS on MAP@100, while multi-channel GRAU alone is initially worse and only later excels in MAP@100.
Limitations
- Gains on Twitter Hits@100 vs. SR+GU are modest; dataset-dependent sensitivity between social and preference graphs.
- Temperature hyperparameter tau must be tuned (best 0.02 reported, 0.05 default).
Experimental setup
Datasets: Twitter Memetracker Douban
Benchmarks: Twitter Memetracker Douban
Metrics: Hits@50 Hits@100 MAP@50 MAP@100
Baselines: DeepDiffuse Topo-LSTM FOREST DyHGCN MS-HGAT CEGCN GRASS
Builds on: GRASS GRAU Structural Spreading GCN hypergraph convolution
Closest papers by meaning
- Enhancing Multi-Scale Diffusion Prediction via Sequential Hypergraphs and Adversarial Learning (AAAI 2024 · similarity 0.85)
- Ghidorah: Towards Robust Multi-Scale Information Diffusion Prediction via Test-Time Training (AAAI 2025 · similarity 0.85)
- Decoupling User Relationships Guides Information Diffusion Prediction (Student Abstract) (AAAI 2024 · similarity 0.84)
- THGNets: Constrained Temporal Hypergraphs and Graph Neural Networks in Hyperbolic Space for Information Diffusion Prediction (AAAI 2025 · similarity 0.84)
- A Thorough Comparison Between Independent Cascade and Susceptible-Infected-Recovered Models (AAAI 2025 · similarity 0.83)
- Towards Effective, Efficient and Unsupervised Social Event Detection in the Hyperbolic Space (AAAI 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.