Jiahe Li

Junior Student in Computer Engineering
Zhejiang University - University of Illinois at Urbana-Champaign Institute

Home

Research Experience

CV

Study

Personal

Social

Hosted on GitHub Pages — Theme by orderedlist

Dynamic Network Embedding

Network embedding, a graph representation learning method illustrating network topology by mapping nodes into lower-dimension vectors, is challenging to accommodate the ever-changing dynamic graphs in practice. We propose the Dynamic Adjacency Matrix Factorization (DAMF) algorithm, which achieves an efficient and accurate dynamic network embedding by rotating and scaling the coordinate system where the network embedding resides with no more than the number of edge modifications changes of node embeddings. Moreover, a dynamic Personalized PageRank is applied to the obtained network embeddings to enhance node embeddings and capture higher-order neighbor information dynamically. We unprecedentedly expand dynamic network embedding experiments to billion-edge graphs, where DAMF updates billion-level parameters in less than 10ms. graph

(Accepted by kdd2023)

Consistency Learning

Epilepsy is one of the most common and serious neurological disorders, affecting approximately 65 million people worldwide. One-third of patients are medication-resistant, and surgical resection of the lesion tissue through neurosurgery is necessary. Stereoelectroencephalography (SEEG) can provide more precise localization of epileptic foci in both the temporal and spatial domains, as it involves implanting electrodes deep into the brain to obtain more comprehensive and stereoscopic information. As the amount of collected data expands, it is crucial to model large-scale data more reasonably and use differential labels. To harmonize training labels from different sources, we propose a hierarchical framework based on consistency learning to address the problem of label consistency.

rORAM + B-tree

Oblivious RAM hides the memory access pattern by using extra bandwidth and memory overhead. Path ORAM stores memory blocks on a binary tree's random branch. Because the blocks are positioned on distinct branches, repeated procedures reveal no information. rORAM is an optimization of Path ORAM to achieve faster interval queries by saving multiple binary trees. Another spatial optimization of Path ORAM is to use B-tree instead of linear position map storage. After studying, I found that by combining the two optimizations, I can take full advantage of both and make progress in both time and space.

GitHub link

Container Vulnerability Reachability

Bloated dependencies are libraries that are packed with the compiled code of an application but are not required to create and run the application. DepClean is a way to automatically clean up a Java project's dependency tree and remove dependencies that are not necessary for building the project. My team members and I learned about this method during our summer research and proposed an optimization that shrinks the unit of dependency study from a file to a function.