CTransE: An Effective Information Credibility Evaluation Method Based on Classified Translating Embedding in Knowledge Graphs

Abstract

With the advent of the Big Data era, new network modes have evolved, such as 5G communication and online social networks, resulting in a dramatically increasing amount of information. However, difficulties occur in traditional information credibility evaluation methods, such as manual analysis. It would cost tremendous manpower and time to distinguish the trusted information from the fictitious ones faced with massive data. Therefore, it is urgent and necessary to come up with a more intelligent method to evaluate the credibility of the information. Aiming at the problem of low information quality and the need of efficient assessment in the big data environment, we present an information credibility evaluation method based on knowledge graphs. Firstly, we propose a CTransE model, a translating embedding model based on the classification optimization, which maps entities and relationships into continuous vector space according to scheduled rules. The method reduces the randomness of the algorithm to enhance the stability and accuracy of vector representation. Secondly, we use parameter adaptive adjustment method to optimize the process of stochastic gradient descent. With this approach, we not only obtain a quick convergence to reduce the time cost, but also acquire a better convergence result of knowledge representation compared with previous methods. Finally, we take both ranking and vector distance into account to calculate the information credibility and feedback the most likely information at the same time. Performance on real datasets shows that average ranking has improved about 4% and accuracy in top ten percent has improved more than 13%. Besides, the method also performs well in the field of knowledge completion, database cleaning and so on. It is a breakthrough for applying knowledge graph to quantitative calculation of information credibility evaluation and the method proves to be effective since extensive experiments show that the performance of CTransE is remarkable superior to previous ones on several large-scale knowledge bases.