生物醫學知識圖譜驗證之特徵工程及知識圖譜嵌入混合法

Abstract

生物醫學知識圖譜在生物醫藥學領域有廣泛的應用，例如提供彙整的知識給醫療從業人員、進行舊藥新用的推論、化學或藥物的發現及藥物間作用的預測等。然而，大型生物醫學知識圖譜（例如 SemMedDB）是從生物醫學文獻中利用自動化的關係萃取技術所得到的，往往充滿錯誤的關聯。若在應用時使用這些含有雜訊的知識圖譜，可能會導致結果表現下降或產生錯誤的結論。 知識圖譜驗證的目標是判斷知識圖譜中實體間的關係是否正確，這項任務可以幫助進行知識圖譜的清理，有效改善其品質。過去知識圖譜驗證的相關研究考量的特徵主要可分為三類：區域特徵、基於知識圖譜嵌入的特徵、與基於路徑的特徵。如 Her (2021)考慮資料的統計資訊產生區域特徵並獲得良好的結果，而知識圖譜嵌入特徵能夠有效地將知識圖譜中的實體與實體關係表達成嵌入表示，另外基於路徑的方法利用專家定義的規則或規則模式來萃取路徑型態的特徵。 本文提出一混合方法，結合以上特徵的特點，使用三類特徵進行生物醫學知識圖譜驗證任務並分析其效能。實驗結果顯示，利用特徵工程產生的統計性區域特徵和利用知識圖譜嵌入產生的全域特徵，均能有效幫助判斷實體間關係的正確性。本文提出的混合方法在準確率、精確度、召回率和 F1 分數方面均優於單純考慮單一特徵的方法，顯示考慮多樣特性的特徵能夠有效提升結果表現。

Biomedical knowledge graphs have extensive applications in the field of biomedicine, such as providing comprehensive knowledge to healthcare practitioners, facilitating drug repurposing, drug discovery, and predicting drug-drug interactions. However, large biomedical knowledge graphs, such as SemMedDB (Kilicoglu et al., 2012), are generated through automated relation extraction techniques from biomedical literature and often contain erroneous associations. Utilizing such noisy knowledge graphs in downstream applications can lead to decreased performance and erroneous conclusions. Knowledge graph verification aims to determine the correctness of relationships between entities in a knowledge graph. This task facilitates knowledge graph cleaning and improves its quality. Previous research on knowledge graph verification has focused on three main types of features: local features, knowledge graph embedding features, and path-based features. For example, Her (2021) considered statistical information to derive local features and achieved promising results. Knowledge graph embedding features effectively represent entity-entity relationships in a focal knowledge graph, while path-based features incorporate expert-defined rules or patterns to extract path-based features for knowledge graph verification purposes. In this thesis, we propose a hybrid approach by combining the strengths of the aforementioned types of features for biomedical knowledge graph verification. We empirically evaluate our proposed approach using a set of knowledge triplets whose correctness is annotated by a domain expert. Experimental results demonstrate that both locally derived statistical features and global features derived from knowledge graph embedding effectively contribute to determining the correctness of entity relationships. This hybrid approach outperforms single-feature-type approaches in accuracy, precision, recall, and F1 score, highlighting the effectiveness of considering diverse features in improving verification effectiveness.