KBGAN: Adversarial Learning for Knowledge Graph Embeddings (1711.04071v3)

Abstract: We introduce KBGAN, an adversarial learning framework to improve the performances of a wide range of existing knowledge graph embedding models. Because knowledge graphs typically only contain positive facts, sampling useful negative training examples is a non-trivial task. Replacing the head or tail entity of a fact with a uniformly randomly selected entity is a conventional method for generating negative facts, but the majority of the generated negative facts can be easily discriminated from positive facts, and will contribute little towards the training. Inspired by generative adversarial networks (GANs), we use one knowledge graph embedding model as a negative sample generator to assist the training of our desired model, which acts as the discriminator in GANs. This framework is independent of the concrete form of generator and discriminator, and therefore can utilize a wide variety of knowledge graph embedding models as its building blocks. In experiments, we adversarially train two translation-based models, TransE and TransD, each with assistance from one of the two probability-based models, DistMult and ComplEx. We evaluate the performances of KBGAN on the link prediction task, using three knowledge base completion datasets: FB15k-237, WN18 and WN18RR. Experimental results show that adversarial training substantially improves the performances of target embedding models under various settings.

• The paper presents an adversarial framework that generates high-quality negative samples to improve knowledge graph embeddings.
• It integrates GAN components with probability-based and translation-based models, demonstrating enhanced link prediction metrics on multiple datasets.
• This method reduces dependency on uniform negative sampling, offering a promising direction for future adversarial approaches in knowledge discovery.

Adversarial Learning for Knowledge Graph Embeddings: An Examination of "kbgan"

The paper "kbgan: Adversarial Learning for Knowledge Graph Embeddings" introduces a novel adversarial learning framework aimed at enhancing the performance of knowledge graph embedding (KGE) models. The authors explore the application of generative adversarial networks (GANs) to generate high-quality negative samples, which is a critical challenge when working with knowledge graphs.

Overview of the Proposed Framework

The kbgan framework leverages a GAN-inspired approach to tackle the problem of generating meaningful negative samples for KGE training. Typically, knowledge graphs such as Freebase or Yago contain only positive instances of data. Conventional methods for creating negative samples, such as replacing entities in triples with random entities, often lead to easily discriminatable negative instances that do not contribute effectively to model training. The kbgan framework addresses this issue by using a GAN structure where the generator creates plausible negative instances, and the discriminator refines its embedding models using these samples.

In kbgan, the generator is designed to output high-quality negative samples by optimizing a probabilistic model, and the discriminator uses these samples to improve its embeddings. The authors focus on a combination of probability-based models (DistMult and ComplEx) as generators and translation-based models (TransE and TransD) as discriminators. This setup allows for flexible application across different KGE models and potentially better performance due to the strategic generation of negative samples.

Experimental Results

The authors conducted experiments using three datasets: FB15k-237, WN18, and WN18RR, primarily evaluating the performance in link prediction tasks. Their approach demonstrated substantial improvements in mean reciprocal ranking (MRR) and hits at 10 (H@10) compared to traditional methods. For instance, the combination of TransE with either DistMult or ComplEx benefited from significant performance gains, indicating the effectiveness of the adversarial setup in generating informative negative samples. This improvement suggests the utility of kbgan in alleviating the weaknesses of uniform negative sampling by providing more challenging examples for the discriminator.

Implications and Future Directions

The kbgan framework presents several theoretical and practical implications. Theoretically, it bridges the gap between adversarial learning and knowledge graph completion, suggesting potential avenues for integrating adversarial methodologies in other aspects of knowledge discovery and representation learning. Practically, kbgan demonstrates a clear pathway to enhance existing KGE methods without the need for additional ontological data, making it feasible to implement in various domains where structured negative data is sparse.

Looking forward, there are intriguing possibilities for future research. The application of this framework could be extended to more complex models beyond the ones tested, potentially leading to state-of-the-art results in various KGE tasks. Furthermore, exploring different architectures for both the generator and discriminator could yield insights into optimal configurations for specific types of knowledge graphs, thereby broadening the applicability and impact of adversarial learning in this field.

Conclusion

The approach detailed in "kbgan: Adversarial Learning for Knowledge Graph Embeddings" showcases a promising solution to the problem of negative sampling in knowledge graph completion. By employing an adversarial learning framework, the authors have effectively improved the quality of negative samples, enhancing the overall performance of KGE models. Their work lays the groundwork for further exploration into adversarial methods within this domain, highlighting the continuing evolution of knowledge representation technologies.