A Re-Valuation of Knowledge Graphs Completion Methods.
When you’re trying to find the number of people in a group, Knowledge Graph Completion (KGC) aims to automatically predict the missing links in large scale.
Few of the major applications are in data mining, machine learning and NLP.
Real world data sets are represented in triage of the world. (h,r,t) ‘r’ represents the relation, ‘h’ and ‘t’ represent the head and the tail spins.
Previous methods have all been using score functions, to measure the plausibility of the triplet. A lot of new neural networks have been proposed, which utilizes the black box neural network, including the conventional neural network (CNNs) Recurrent (RNN) and Graphical (GNN).
The biggest disadvantage is the results of the Neural Network are not consistent across datasets — and the missing links in the output are not analysed.
There are many evaluation processes which do not analyse the models extensively. The “Knowledge Graph Completion” evaluation can be analysed in 2 methods: Non-Affected and Affected.
Non-affected gives consistent results across different protocols. The ‘Affected’ consists of recent Artificial Neural Network, where performance is affected by different evaluation protocols.
All these protocols can be grouped into 3: “TOP, RANDOM, BOTTOM”
TOP and BOTTOM evaluation protocol has its own biases, which it comes from different datasets.
Based on the observations, the proposition is RANDOM evaluation protocol that can clearly distinguish between these affected methods from others. We strongly recommend the research community to follow the RANDOM evaluation protocol for all KGC evaluation purposes.
