On 17 October, VNUHCM–University of Science (HCMUS) held a university-level doctoral thesis defence for Nguyễn Ngọc Trung, a doctoral researcher in Computer Science, whose thesis was entitled “Event Prediction in Temporal Knowledge Graphs”. The thesis was supervised by Prof. Bùi Xuân Hải (principal supervisor) and Dr Bùi Xuân Bình Minh (co-supervisor). The examination committee comprised Assoc. Prof. Võ Thị Ngọc Châu (first examiner), Dr Nguyễn Tiến Huy (second examiner) and Assoc. Prof. Phạm Duy Khánh (third examiner).

A Temporal Knowledge Graph (TKG) represents knowledge by associating temporal information with relationships and events, thereby modelling how knowledge evolves over time. The thesis investigated event prediction in temporal knowledge graphs across two main tasks: interpolation, which aims to complete missing events or links within an observed period, and extrapolation, which seeks to forecast events that may occur in the future.

To address these challenges, Doctoral Researcher Nguyễn Ngọc Trung proposed three representation learning and reasoning models designed to improve the extraction of temporal information and event prediction from historical data.

The TBicomR model was developed in the double-complex number space, using rotation-based operations to represent entities, relations and temporal information within knowledge graphs. By combining circular and hyperbolic rotations, the model can represent various relation patterns, including symmetric, antisymmetric, inverse and temporally periodic relations. The model achieves linear time complexity, (O(d)), where (d) denotes the embedding dimension. Performance was evaluated on the benchmark datasets ICEWS14, ICEWS05-15, GDELT, YAGO11k and Wikidata12k.

The FTPComplEx model extends tensor decomposition-based temporal knowledge graph embedding by introducing a flexible temporal weighting mechanism controlled by two parameters. This mechanism allows the influence of temporal information to be adjusted according to the characteristics of each dataset, while maintaining computational complexity comparable to the baseline TPComplEx model. On the YAGO15k dataset, FTPComplEx improved the Mean Reciprocal Rank (MRR) score by approximately 9% and Hits@1 by approximately 11% compared with the baseline under the same evaluation setting.

The HGCT (History Graph Convolution Transformer) model was proposed for extrapolation, combining Transformer architecture with graph convolution to exploit historical event information for future prediction. The model integrates Fact Graph Transformer to capture historical graph structures, Temporal Convolution Transformer to encode temporal information and capture short- and long-term dependencies, and the Query-ConvTransE decoder to predict events. Across six benchmark datasets — ICEWS14, ICEWS18, ICEWS05-15, GDELT, YAGO and WIKI — HGCT improved Hits@k scores by approximately 4.33% and MRR by approximately 3.84% compared with the state-of-the-art baseline used for comparison, under a time-aware filtered evaluation setting.

The models were evaluated using widely adopted metrics in machine learning and knowledge graph research, including Mean Reciprocal Rank (MRR), Hits@k and Mean Absolute Error (MAE). The findings contribute to the development of methods for temporal knowledge graph representation and event prediction, while opening up potential applications in fields involving time-varying data, including finance, healthcare, social networks, e-commerce and cybersecurity.

Alongside these findings, the thesis identified several directions for future research, including handling large-scale, noisy and incomplete datasets; developing representation learning methods for continuous-time environments; improving model interpretability; optimising computational costs; and integrating multimodal data with large language models.

The research findings have been published in three papers in Q1-ranked international journals, including Knowledge-Based Systems and Engineering Applications of Artificial Intelligence, as well as a paper presented at the international conference ESANN 2024.

The examination committee commended the scientific rigour, originality and practical relevance of the research, and unanimously approved Doctoral Researcher Nguyễn Ngọc Trung’s thesis defence.

The full thesis is available for consultation at the Library of VNUHCM–University of Science, the Viet Nam National University Ho Chi Minh City Library and the General Sciences Library of Ho Chi Minh City.