Duc Q. Nguyen

From the end of 2019, one of the most serious and largest spread pandemics occurred in Wuhan (China) named Coronavirus (COVID-19). As reported by the World Health Organization, there are currently more than 100 million infectious cases with an average mortality rate of about five percent all over the world. To avoid serious consequences on people’s lives and the economy, policies and actions need to be suitably made in time. To do that, the authorities need to know the future trend in the development process of this pandemic. This is the reason why forecasting models play an important role in controlling the pandemic situation. However, the behavior of this pandemic is extremely complicated and difficult to be analyzed, so that an effective model is not only considered on accurate forecasting results but also the explainable capability for human experts to take action pro-actively. With the recent advancement of Artificial Intelligence (AI) techniques, the emerging Deep Learning (DL) models have been proving highly effective when forecasting this pandemic future from the huge historical data. However, the main weakness of DL models is lacking the explanation capabilities. To overcome this limitation, we introduce a novel combination of the Susceptible-Infectious-Recovered-Deceased (SIRD) compartmental model and Variational Autoencoder (VAE) neural network known as BeCaked. With pandemic data provided by the Johns Hopkins University Center for Systems Science and Engineering, our model achieves 0.98 \\R^2\\and 0.012 MAPE at world level with 31-step forecast and up to 0.99 \\R^2\\and 0.0026 MAPE at country level with 15-step forecast on predicting daily infectious cases. Not only enjoying high accuracy, but BeCaked also offers useful justifications for its results based on the parameters of the SIRD model. Therefore, BeCaked can be used as a reference for authorities or medical experts to make on time right decisions.

For humans, the COVID-19 pandemic and Coronavirus have undeniably been a nightmare. Although there are effective vaccines, specific drugs are still urgent. Normally, to identify potential drugs, one needs to design and then test interactions between the drug and the virus in an in silico manner for determining candidates. This Drug-Target Interaction (DTI) process, can be done by molecular docking, which is too complicated and time-consuming for manual works. Therefore, it opens room for applying Artificial Intelligence (AI) techniques. In particular, Graph Neural Network (GNN) attracts recent attention since its high suitability for the nature of drug compounds and virus proteins. However, to introduce such a representation well-reflecting biological structures of biological compounds is not a trivial task. Moreover, since available datasets of Coronavirus are still not highly popular, the recently developed GNNs have been suffering from overfitting on them. We then address those issues by proposing a novel model known as Atom-enhanced Graph Neural Network with Multi-hop Gating Mechanism. On one hand, our model can learn more precise features of compounds and proteins. On the other hand, we introduce a new gating mechanism to create better atom representation from non-neighbor information. Once applying transfer learning from very large databanks, our model enjoys promising performance, especially when experimenting with Coronavirus.

Recent advancements in large language models (LLMs) have underscored their importance in the evolution of artificial intelligence. However, despite extensive pretraining on multilingual datasets, available open-sourced LLMs exhibit limited effectiveness in processing Vietnamese. The challenge is exacerbated by the absence of systematic benchmark datasets and metrics tailored for Vietnamese LLM evaluation. To mitigate these issues, we have finetuned LLMs specifically for Vietnamese and developed a comprehensive evaluation framework encompassing 10 tasks and 31 metrics. We observe that finetuning can help LLMs transfer knowledge across languages, serving as an efficient way to bolster their capabilities in non-English languages. Moreover, our analysis indicates that larger models can introduce more biases and uncalibrated outputs and the key factor influencing LLM performance is the quality of the training or finetuning datasets. These insights underscore the significance of meticulous finetuning with high-quality datasets in enhancing LLM performance.

Subgraph matching is a challenging problem with a wide range of applications in drug discovery, social network analysis, biochemistry, and cognitive science. It involves determining whether a given query graph is present within a larger target graph. Traditional graph-matching algorithms provide precise results but face challenges in large graph instances due to the NP-complete nature of the problem, limiting their practical applicability. In contrast, recent neural network-based approximations offer more scalable solutions but often lack interpretable node correspondences. To address these limitations, this article presents a multi-task learning framework called xNeuSM: Explainable Neural Subgraph Matching, which introduces Graph Learnable Multi-hop Attention Networks (GLeMA) that adaptively learn the parameters governing the attention factor decay for each node across hops rather than relying on fixed hyperparameters. Our framework jointly optimizes both subgraph matching and finding subgraph-isomorphism mappings. We provide a theoretical analysis establishing error bounds for GLeMA’s approximation of multi-hop attention as a function of the number of hops. Additionally, we prove that learning distinct attention decay factors for each node leads to a correct approximation of multi-hop attention. Empirical evaluation on real-world datasets shows that xNeuSM achieves substantial improvements in prediction F1 score of up to 34% compared to approximate baselines and, notably, at least a seven-fold faster query time than exact algorithms. With these results, xNeuSM can be applied to solve matching problems in various domains spanning from biochemistry to social science.

Bayesian optimization (BO) is a popular framework for optimizing black-box functions, leveraging probabilistic models such as Gaussian processes. Conventional BO algorithms, however, assume static query costs, which limit their applicability to real-world problems with dynamic cost structures such as geological surveys or biological sequence design, where query costs vary based on the previous actions. We propose a novel nonmyopic BO algorithm named LookaHES featuring dynamic cost models to address this. LookaHES employs a neural network policy for variational optimization over multi-step lookahead horizons to enable planning under dynamic cost environments. Empirically, we benchmark LookaHES on synthetic functions exhibiting varied dynamic cost structures. We subsequently apply LookaHES to a real-world application in protein sequence design using a large language model policy, demonstrating its scalability and effectiveness in handling multi-step planning in a large and complex query space. LookaHES consistently outperforms its myopic counterparts in synthetic and real-world settings, significantly improving efficiency and solution quality. Our implementation is available at https://github.com/sangttruong/nonmyopia.