Are Message Passing Neural Networks Really Helpful for Knowledge Graph Completion?

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 10696–10711 July 9-14, 2023 ©2023 Association for Computational Linguistics Are Message Passing Neural Networks Really Helpful for Knowledge Graph Completion? Juanhui Li1, Harry Shomer1 , Jiayuan Ding1 , Yiqi Wang1∗, Yao Ma2 Neil Shah3, Jiliang Tang1, Dawei Yin4 1Michigan State University, 2New Jersey Institute of Technology 3Snap Inc., 4Baidu Inc. {lijuanh1,shomerha,dingjia5,wangy206,tangjili}@msu.edu yao.ma@njit.edu, nshah@snap.com, yindawei@acm.org Abstract Knowledge graphs (KGs) facilitate a wide va- riety of applications. Despite great efforts in creation and maintenance, even the largest KGs are far from complete. Hence, KG comple- tion (KGC) has become one of the most cru- cial tasks for KG research. Recently, consider- able literature in this space has centered around the use of Message Passing (Graph) Neural Networks (MPNNs), to learn powerful embed- dings. The success of these methods is nat- urally attributed to the use of MPNNs over simpler multi-layer perceptron (MLP) models, given their additional message passing (MP) component. In this work, we find that surpris- ingly, simple MLP models are able to achieve comparable performance to MPNNs, suggest- ing that MP may not be as crucial as previ- ously believed. With further exploration, we show careful scoring function and loss func- tion design has a much stronger influence on KGC model performance. This suggests a conflation of scoring function design, loss function design, and MP in prior work, with promising insights regarding the scalability of state-of-the-art KGC methods today, as well as careful attention to more suitable MP de- signs for KGC tasks tomorrow. Our codes are publicly available at: https://github.com/ Juanhui28/Are_MPNNs_helpful. 1 Introduction Knowledge graphs (KGs) (Bollacker et al., 2008; Carlson et al., 2010) are a type of knowledge base, which store multi-relational factual knowledge in the form of triplets. Each triplet specifies the re- lation between a head and a tail entity. KGs con- veniently capture rich structured knowledge about many types of entities (e.g. objects, events, con- cepts) and thus facilitate numerous applications such as information retrieval (Xiong et al., 2017a), recommender systems (Wang et al., 2019), and ∗Corresponding Author question answering (West et al., 2014). To this end, the adopted KGs are expected to be as comprehen- sive as possible to provide all kinds of required knowledge. However, existing large-scale KGs are known to be far from complete with large portions of triplets missing (Bollacker et al., 2008; Carlson et al., 2010). Imputing these missing triplets is of great importance. Furthermore, new knowledge (triplets) is constantly emerging even between ex- isting entities, which also calls for dedicated efforts to predict these new triplets (García-Durán et al., 2018; Jin et al., 2019). Therefore, knowledge graph completion (KGC) is a problem of paramount im- portance (Lin et al., 2015; Yu et al., 2021). A cru- cial step towards better KGC performance is to learn low-dimensional continuous embeddings for both entities and relations (Bordes et al., 2013). Recently, due to the intrinsic graph-structure of KGs, Graph Neural Networks (GNNs) have been adopted …

Content truncated to 500 words