Research Projects

Advisor: Prof. Xiaofang ZHOU - Hong Kong University of Science and Technology, HK

Discrete spatial-temporal data is not easy to handle because of some challenging properties, such as multi- dimensionality and large-scale.
We aim to design and implement an integrated point cloud index by leveraging the original index method as a foundation. Our ultimate objective is to realize the design of a database index system that supports time queries while ensuring stable and efficient query performance.
Owing to the unbalanced structure of the existing method, the time cost for queries becomes unstable and varies with different data distributions. The initial solution proposed by the project is to employ a balanced search tree.

Advisor: Dr. Yangqiu Song - Hong Kong University of Science and Technology, HK

With the increasing availability of embedding vector databases, the potential for recovering the original data from embeddings has emerged as a critical privacy concern. When data owners send embeddings to a third party that is not entirely trusted, there is a risk of the third party potentially victimizing the embeddings through inversion techniques to reconstruct the original data.
Under an iterative method of generative inversion, employing a generative model for the reconstruction of original data has proven to be effective. We propose a novel architecture that utilizes an iterative method to reconstruct both node and relation information. The iterative mechanism aids in reinforcing and refining the recovery results at each iteration, leading to improved accuracy.

Advisor: Dr. Yangqiu Song - Hong Kong University of Science and Technology, HK

The performance of traditional embedding methods, which do not specifically consider numbers, has been inade- quate when applied to number-sensitive knowledge graphs.
We analyzed and improve the current knowledge graph embedding method by utilizing of more sophisticated knowledge graph embedding techniques specifically tailored for numbers.
The experiments demonstrated that incorporating spatial considerations of numbers significantly enhances the performance of the task. Number embedding can be instrumental in facilitating classification tasks on datasets that are sensitive to numerical information. However, significant differences persist due to variations in the numerical range of the data.

Google has substantiated that incorporating a “theory of mind” prompt can be advantageous for large language models in generating intension. The project introduces a multi-step approach using the theory of mind prompt to effectively guide the generation of shopping intentions, thereby offering a novel method for facilitating online shopping decisions.