Generalized Dot Product Models for Link Prediction

Identifier:  TFG:9482

Handle: https://hdl.handle.net/20.500.11797/TFG9482

Authors:  Llobera Querol, Irene

Abstract:
Understanding network structure is key across disciplines using graphs. Two common models are the Random Dot Product (RDP) and the Mixed Membership Stochastic Block Model (MMSBM). RDP is simple and interpretable but struggles with disassortative networks. We analyze this limitation, provide empirical evidence, and show MMSBM’s greater flexibility. To address RDP’s shortcoming, we introduce a complex-valued generalization that allows negative inner products, enabling it to model disassortative structures. Through experiments on synthetic and real-world networks, we demonstrate that while standard RDP fails, both MMSBM and our new model successfully capture disassortative patterns.