Вісник № 01. Системний аналіз, управління та інформаційні технології
Постійне посилання колекціїhttps://repository.kpi.kharkov.ua/handle/KhPI-Press/42344
Переглянути
Документ Using graph embeddings for Wikipedia link prediction(НТУ "ХПІ", 2019) Shaptala, Roman Vitaliyovych; Kyselev, Gennadiy DmytrovychLink prediction is an important area of study in network analysis and graph theory which tries to answer the question of whether two nodes in the graph might have an association in the future. Nowadays, graphs are ubiquitously present in our lives (social networks, circuits, roads etc.), which is why the problem is crucial to the development of intelligent applications. In the past, there have been proposed methods of solving link prediction problem through algebraic formulations and heuristics, however, their expressive power and transferability fell short. Recently, graph embedding methods have risen to popularity because of their effectiveness and the ability to transfer knowledge between tasks. Inspired by the famous in machine learning and natural language processing research Word2Vec approach, these methods try to learn a distributed vector representation, called an embedding, of graph nodes. After that a binary classifier given a pair of embeddings predicts the probability of the existence of a link between the encoded nodes. In this paper, we review several graph embedding approaches for the problem of Wikipedia link prediction, namely Wikipedia2vec, Role2vec, AttentionWalk and Walkets. Wikipedia link prediction tries to find pages that should be interlinked due to some semantic relation. We evaluate prediction accuracy on a hold-out set of links and show which one proves to be better at mining associations between Wikipedia concepts. The results include qualitative (principal component analysis dimensionality reduction and visualization) and quantitative (accuracy) differences between the proposed methods. As a part of the conclusion, further research questions are provided, including new embedding architectures and the creation of a graph embedding algorithms benchmark.