Research
My research interests are in the following areas:
Data Mining, Machine Learning, and Deep Neural Networks Algorithms
Business Data Science that applies AI/ML to business domains: finance, FinTech, biomedical, and healthcare.
I have several grants to sponsor my research on using ML in healthcare and biomedical data analysis, FinTech, and scientific data analysis. If you are interested in research topics in healthcare, EHR, FinTech (Bond, Stock, and Mutual Funds), please send me message.
Business Data Science
Use Graph Neural Network to process spatial-temporal data from the global climate model, power grid, traffic prediction, and distributed sensors. We first apply the attention mechanism to connect the "dots" (sensors, monitoring devices) and learn dynamic network structures among the system components over time. Next, the end-to-end graph neural networks pipeline diffuses and propagates the sensors' reading and configuration settings into the learned networks, ultimately predicting the future status. We will also design supervised ML models to uncover the intrinsic relationship between the overall network and the system performance, events, and anomalies.
Many recent machine learning algorithms were applied to financial data and gained some traction. In this research, we treat finance that has spatial-temporal relationships, i.e., assets are similar to other assets (spatial) and have cyclic patterns (temporal). Graph neural network, combined with recurrent neural networks,
provides effective solutions on asset prices, market prediction, and factor discovery.
Complex Network Models for Asset pricing and Financial Decision Making: FinTech Via A Network Len
Asset pricing plays a vital role in financial investment decisions and still relies on the linear Fama-French models to identify risk factors and estimate excessive returns. AI and machine learning-based asset pricing models are still in their early stage but demonstrates their effectiveness in improving decision-making. Recent studies suggest that networks among firms (sectors) play an essential role in asset valuation. It is challenging to capture and investigate the implications incurred by those networks because of the continuous evolution of networks in response to market micro and macro changes. To address this challenge, we propose to develop an end-to-end graph neural network model and show its applicability in asset pricing. First, we apply the attention mechanism to learn dynamic network structures of the equity market over time and then use a recurrent convolutional neural network to diffuse and propagate firms' features in the learned networks.
Co-investment networks
Multiple networks (causality, supply-chain, co-investment, and board member networks) exist among firms and provide a rich set of contexts for detecting global and local trends. The asset pricing might be affected by multiple types of networks simultaneously. To cope with these challenges, we apply the self-supervised approach to learn a consensus graph by exploring the graph embeddings of node features concerning multiple graphs (multi-view). We adopt state-of-the-art graph neural networks (Eigenlearn, APPNP) to generate graph embeddings from individual graph views. The learning objective has two terms: the universal smoothness term in the learned graph and a contrastive loss that regularizes the learned graph to distinguish the nearest neighbors (positive samples) in each graph context from those distant (negative samples). It is necessary to perform graph augmentation to ensure model stability, such as add/drop edges and mask node features. The superiority of our model will be demonstrated through accurate return prediction and enhanced portfolio Sharpe ratio. The dynamic network learned from our model will model general market conditions over time and explain the equity network systemic risks previously identified in the literature.
Utilize Graph neural network models to Uncover the fundamental of our Universe
Trigger (interesting events) detection is crucial to high-energy and nuclear physics experiments because it improves data acquisition efficiency. It also plays a vital role in facilitating the downstream offline data analysis process. The sPHENIX detector, located at the Relativistic Heavy Ion Collider in Brookhaven National Laboratory, is one of the largest nuclear physics experiments on a world scale and is optimized to detect physics processes involving charm and beauty quarks. The formation of the early universe can be better understood by studying particles produced from collisions between two proton beams, two gold nuclei beams, or a mixture of both. This paper presents a model architecture for trigger detection with geometric information from two fast silicon detectors. Transverse momentum is introduced as an intermediate feature from physics heuristics. We also prove its importance through our training experiments. Each event consists of tracks and can be viewed as a graph. A bipartite graph neural network is integrated with the attention mechanism to design a binary classification model. Compared with the state-of-the-art algorithm for trigger detection, our model is parsimonious and increases the accuracy and the AUC score by more than 15%.
Graph neural network models
We will design new graph kernels based on the Fisher information matrix, heat diffusion, and Green's functions. We learn and construct graphs based on aggregated heat kernel, which has proved to be a much better performance and fault tolerance than traditional spectral techniques. The aggregated heat kernels are evaluated by the computation of eigenvalues and eigenvectors. This unavoidably involves solving the eigenvalues and eigenvectors problem of large-scale matrices. We can also explore the Krylov algorithm to extract approximate Eigenvalues and eigenvectors for extensive graphs.
Residual learning and Eigenvalue perturbation in Graph Neural networks. Network structured data is ubiquitous in natural and social science applications. Graph Convolutional Neural Network (GCN) has attracted significant attention recently due to its success in representing, modeling, and predicting large-scale network data. Various graph convolutional filters were proposed to process graph signals to boost graph-based semi-supervised learning. We will introduce a novel spectral learning technique called EigLearn, which uses residual learning to perturb the graph filter matrix's eigenvalues to optimize its capability. EigLearn is relatively easy to implement, yet thorough experimental studies reveal that it is more effective and efficient than the prior works on the specific issue. EigLearn only perturbs a small number of eigenvalues and does not require a complete eigendecomposition.