Scalable Binary Representation Learning and Proactive Control for Adversarial Information Spread in Networks via Optimization

Abstract

Network-structured data arise in social platforms, knowledge bases, and large-scale information systems. As these networks grow in size and complexity, scalable representation learning becomes essential for efficient retrieval, reasoning, and analytics. While continuous embedding methods have been widely explored, their reliance on high-dimensional real-valued vectors leads to substantial memory usage and limits scalability in large systems. At the same time, real-world networks frequently exhibit adversarial diffusion processes—such as misinformation spread—highlighting the need not only to represent network structure efficiently but also to make timely and strategic decisions on the network. This dissertation advances a unified research agenda centered on two complementary directions: (i) scalable binary representation learning for complex networks, and (ii) optimization-based proactive control of adversarial information spread. The first direction is to develop binary learning frameworks for information networks and knowledge graphs. For information networks, we introduce a matrix-factorization-based hashing method with margin regulation, thresholding for robustness, and then develop a MaxCut reformulation enabling efficient discrete optimization. For knowledge graphs, we propose a fully discrete embedding framework that models relation-specific transformations as element-wise rotations, supported by an alternating optimization algorithm with dynamic range adjustment. Experiments across multiple datasets demonstrate that both methods significantly improve retrieval efficiency and embedding quality over state-of-the-art techniques. The second direction is to address adversarial diffusion and defense. We introduce a detection-driven Attacker–Monitor–Defender model and formulate the defender’s planning problem as a robust Stackelberg optimization. A simulation-based evaluation under the Independent Cascade model enables effective strategy assessment. Experiments show that detection-guided and robustness-aware interventions substantially reduce adversarial impact compared with existing methods. Together, these contributions provide scalable binary learning tools and principled optimization-based intervention strategies, advancing both the theoretical foundations and practical methodologies for intelligent computation on large networked systems.