Article Source
Dynamic Graph Information Bottleneck
- Temporal Graph Learning Reading Group
- Paper: βDynamic Graph Information Bottleneckβ
- Speaker: Haonan Yuan
- Date: March 21, 2024
Abstract
Dynamic Graphs widely exist in the real world, which carry complicated spatial and temporal feature patterns, challenging their representation learning. Dynamic Graph Neural Networks (DGNNs) have shown impressive predictive abilities by exploiting the intrinsic dynamics. However, DGNNs exhibit limited robustness, prone to adversarial attacks. This paper presents the novel Dynamic Graph Information Bottleneck (DGIB) framework to learn robust and discriminative representations. Leveraged by the Information Bottleneck (IB) principle, we first propose the expected optimal representations should satisfy the Minimal-Sufficient-Consensual (MSC) Condition. To compress redundant as well as conserve meritorious information into latent representation, DGIB iteratively directs and refines the structural and feature information flow passing through graph snapshots. To meet the MSC Condition, we decompose the overall IB objectives into DGIBππ and DGIBπΆ, in which the DGIBππ channel aims to learn the minimal and sufficient representations, with the DGIBπΆ channel guarantees the predictive consensus. Extensive experiments on real-world and synthetic dynamic graph datasets demonstrate the superior robustness of DGIB against adversarial attacks compared with state-of-the-art baselines in the link prediction task. To the best of our knowledge, DGIB is the first work to learn robust representations of dynamic graphs grounded in the information-theoretic IB principle.
λμ κ·Έλνμ κ²¬κ³ ν νν νμ΅μ μν λμ κ·Έλν μ 보 λ³λͺ© (DGIB) νλ μμν¬
μ€ μΈκ³μλ λ€μν 곡κ°μ , μκ°μ νΉμ§ ν¨ν΄μ κ°μ§λ λμ κ·Έλνκ° λ리 μ‘΄μ¬νλ©°, μ΄λ κ·Έλν νν νμ΅μ μμ΄ μ΄λ €μμ μ μν©λλ€. λμ κ·Έλν μ κ²½λ§ (DGNN)μ λ΄μ¬μ λμ νΉμ±μ νμ©νμ¬ μΈμμ μΈ μμΈ‘ λ₯λ ₯μ 보μ¬μ£Όμμ§λ§, μ λμ 곡격μ μ·¨μ½νλ€λ νκ³κ° μμ΅λλ€. μ΄ λ Όλ¬Έμμλ κ²¬κ³ νκ³ νλ³λ ₯ μλ ννμ νμ΅νκΈ° μν μλ‘μ΄ λμ κ·Έλν μ 보 λ³λͺ© (DGIB) νλ μμν¬λ₯Ό μ μν©λλ€.
μ 보 λ³λͺ© (IB) μ리 νμ©μ ν΅ν΄ λ¨Όμ , κΈ°λ μ΅μ ννμ μ΅μ-μΆ©λΆ-ν©μ (MSC) 쑰건μ λ§μ‘±ν΄μΌ νλ€κ³ μ μν©λλ€. DGIBλ μ 보 λ³λͺ© μ리λ₯Ό μ΄μ©νμ¬ λΆνμν μ 보λ₯Ό μμΆνκ³ μ€μν μ 보λ₯Ό μ μ¬μ ννμΌλ‘ μ μ§νκΈ° μν΄ κ·Έλν μ€λ μ·μ ν΅κ³Όνλ ꡬ쑰μ λ° νΉμ§ μ 보 νλ¦μ λ°λ³΅μ μΌλ‘ μ§μ λ° κ°μ ν©λλ€. MSC 쑰건μ μΆ©μ‘±νκΈ° μν΄ μ 체 IB λͺ©μ ν¨μλ₯Ό DGIB_MS λ° DGIB_Cλ‘ λΆν΄ν©λλ€. μ¬κΈ°μ DGIB_MS μ±λμ μ΅μνμ μΆ©λΆν ννμ νμ΅νλ κ²μ λͺ©νλ‘ νκ³ , DGIB_C μ±λμ μμΈ‘ ν©μλ₯Ό 보μ₯ν©λλ€.
μ€μ λ° ν©μ± λμ κ·Έλν λ°μ΄ν° μΈνΈμ λν κ΄λ²μν μ€νμ ν΅ν΄ μ΅μ²¨λ¨ κΈ° baseline λͺ¨λΈκ³Ό λΉκ΅νμ¬ λ§ν¬ μμΈ‘ μμ μμ DGIBμ κ²¬κ³ μ±μ΄ λ°μ΄λλ€λ κ²μ μ μ¦νμ΅λλ€. μ μλ€μ μ°κ΅¬ κ²°κ³Όμ λ°λ₯΄λ©΄, DGIBλ μ 보 μ΄λ‘ μ IB μ리 κΈ°λ°μΌλ‘ λμ κ·Έλνμ κ²¬κ³ ν ννμ νμ΅νλ μ΅μ΄μ μ°κ΅¬μ λλ€.