NATIVE: Multi-modal Knowledge Graph Completion in the Wild This paper proposes a novel framework called NATIVE for multi-modal knowledge graph completion (MMKGC) in the wild. NATIVE addresses the challenges of diversity and imbalance in multi-modal knowledge graphs (MMKGs) by incorporating adaptive fusion and adversarial training. The framework consists of two key modules: Relation-guided Dual Adaptive Fusion (ReDAF) and Collaborative Modality Adversarial Training (CoMAT). ReDAF enables adaptive fusion of any modality with relation guidance, while CoMAT enhances imbalanced modality information through adversarial training. The framework is evaluated on a new benchmark called WildKGC with five datasets, achieving state-of-the-art performance across different scenarios. The empirical results show that NATIVE outperforms existing baselines in terms of accuracy and efficiency. The framework is also theoretically analyzed to justify its design. NATIVE is capable of handling various modalities, including numeric, audio, text, image, and video, and is designed to be generalizable and efficient. The framework is evaluated on multiple datasets and shows strong performance in link prediction tasks. The results demonstrate that NATIVE is effective in handling imbalanced modality information and can generalize to different scenarios. The framework is also compared with other MMKGC methods and shows superior performance in terms of accuracy and efficiency. The paper concludes that NATIVE is a promising approach for MMKGC in the wild.NATIVE: Multi-modal Knowledge Graph Completion in the Wild This paper proposes a novel framework called NATIVE for multi-modal knowledge graph completion (MMKGC) in the wild. NATIVE addresses the challenges of diversity and imbalance in multi-modal knowledge graphs (MMKGs) by incorporating adaptive fusion and adversarial training. The framework consists of two key modules: Relation-guided Dual Adaptive Fusion (ReDAF) and Collaborative Modality Adversarial Training (CoMAT). ReDAF enables adaptive fusion of any modality with relation guidance, while CoMAT enhances imbalanced modality information through adversarial training. The framework is evaluated on a new benchmark called WildKGC with five datasets, achieving state-of-the-art performance across different scenarios. The empirical results show that NATIVE outperforms existing baselines in terms of accuracy and efficiency. The framework is also theoretically analyzed to justify its design. NATIVE is capable of handling various modalities, including numeric, audio, text, image, and video, and is designed to be generalizable and efficient. The framework is evaluated on multiple datasets and shows strong performance in link prediction tasks. The results demonstrate that NATIVE is effective in handling imbalanced modality information and can generalize to different scenarios. The framework is also compared with other MMKGC methods and shows superior performance in terms of accuracy and efficiency. The paper concludes that NATIVE is a promising approach for MMKGC in the wild.