Abstract: Knowledge recombination is pivotal for fostering innovation and interdisciplinary integration. Existing studies typically rely on homogenous knowledge networks for its early prediction, which fail to capture the intricate relationships between knowledge units and their associated entities, thereby constraining predictive performance. To address this limitation, this paper proposes a knowledge-recombination prediction framework based on heterogeneous graph neural networks. The framework integrates multiple heterogeneous entities and relations closely related to knowledge units, constructs an enriched heterogeneous knowledge network through diverse connection strategies, and employs a relation-aware graph convolutional network to predict potential recombination links. Empirical experiments in the cancer immunotherapy domain demonstrate that the proposed framework markedly outperforms traditional homogenous-network baselines, with the F1 score rising from 0.706 to 0.889. The results also confirm that connection strategies for heterogeneous nodes significantly influence predictive performance, underscoring the importance of heterogenous network design in knowledge-recombination prediction.

Key words: Knowledge recombination prediction, Heterogeneous graph neural network, Link prediction, Complex networks, Multiple relationship