Hello, everyone. My name is Zhikai Chen. Currently, I am a third-year Ph.D. student at Michigan State University. Previously, I received both my bachelor's and master's degrees from Shanghai Jiao Tong University. My research primarily focuses on automating machine learning for large-scale industrial relational data, *encompassing automatic machine learning*, *foundation model development*, and *theoretical understanding*. ## Selected Publications ### AutoML AutoG: Towards automatic graph construction from tabular data **Zhikai Chen**, Han Xie, Jian Zhang, Xiang Song, Jiliang Tang, Huzefa Rangwala, George Karypis; ICLR 2025 (poster) [Code](https://github.com/amazon-science/Automatic-Table-to-Graph-Generation) > [!faq]- Description > This paper tries to address a fundamental problem when applying GML to industrial data: How to come up with a downstream-task-friendly graph structure? By utilizing an LLM-agent-based pipeline with a constrained action space, AutoG can effectively construct a graph on non-relational tables extracted from data lakes. Example schemas applied to [rel-f1](https://relbench.stanford.edu/datasets/rel-f1/). Beyond the datasets used in this paper, AutoG can also enhance schemas in relbench and improve downstream task performance. ![[Pasted image 20250619202359.png]] ### Foundation model for relational data Exploring the Potential of Large Language Models (LLMs) in Learning on Graphs [preprint version](https://arxiv.org/abs/2307.03393v3) [journal version](https://arxiv.org/abs/2307.03393), **Zhikai Chen**, Haitao Mao, Hang Li, Wei Jin, Hongzhi Wen, Xiaochi Wei, Shuaiqiang Wang, Dawei Yin, Wenqi Fan, Hui Liu, Jiliang Tang 2023 [Code](https://github.com/CurryTang/Graph-LLM); SIGKDD Explorations and NeurIPS GLFrontiers 2023 > [!faq]- Description > This project is a very early-stage exploration on the usage of large language models for graph machine learning problems, with a focus on node classification task. The key findings include > * We, for the first-time, propose two paradigms to utilize LLMs for graph ML problems, which is LLM-as-predictor and LLM-as-enhancer. > * Many past GNN designs focus on designing different message passing mechanisms, however, we find that utilizing the embedding of Sentence-transformers, an overlooked embedding model, vanilla GNNs can achieve promising performance on datasets like OGB-Arxiv, this inspires future research like OneForAll. > * The scaling law of language models doesn't present when we adopt them as an embedding model for node features, larger models don't necessarily give better performance. > * LLM presents superior zero-shot prediction performance with text-formatted input, this inspires the future application of applying them as a pseudo annotator. At the same time, we find that many answers from LLMs are incorrect while reasonable, which inspires future research on OOD and label bias for node classification tasks. > LLM shows superior performance when dealing with out-of-distribution samples ![[Pasted image 20250624203325.png]] [Label-free Node Classification on Graphs with Large Language Models (LLMS)](https://arxiv.org/abs/2310.04668), **Zhikai Chen**, Haitao Mao, Hongzhi Wen, Haoyu Han, Wei Jin, Haiyang Zhang, Hui Liu, Jiliang Tang [Code](https://github.com/CurryTang/LLMGNN); ICLR 2024(poster) > [!faq]- Description > Large language models present extraordinary effectiveness when generating pseudo labels. However, their efficiency is still limited considering the inference latency. As a result, in this project, we explore its potential as a pseudo annotator, and then utilize message passing to propagate these pseudo labels towards the whole graph. > The key research question lies in 1. selecting a set of samples with good annotation quality; 2. selecting a set of samples that present larger influence. With these two perquisites, we can achieve better prediction quality with lower annotation budgets. ![[Pasted image 20250627103406.png]] > We then come up with a pipeline combining the importance of queried nodes together with an approximation of the annotation quality. > We check the effectiveness of our method on node classification tasks ranging from small-scale to large-scale datasets, and the improvement is consistent ![[Pasted image 20250627152654.png]] [Graph Foundation Models](https://arxiv.org/abs/2402.02216) Haitao Mao\*, **Zhikai Chen\***, Wenzhuo Tang, Jianan Zhao, Yao Ma, Tong Zhao, Neil Shah, Michael Galkin, Jiliang Tang 2024 [[Paper lists](https://github.com/CurryTang/Towards-Graph-Foundation-Models-New-perspective-)]; ICML 2024 (Spotlight) > [!faq]- Description > Graph Foundation Models (GFMs) represent a nascent yet rapidly evolving frontier in graph learning, aiming to leverage large-scale, diverse graph data for broad downstream applicability rather than training task-specific GNNs from scratch. This paper argues that the key to successful GFMs lies in constructing a “graph vocabulary”—the basic transferable units that encode invariances across heterogeneous graphs—which it grounds in principles from network analysis, expressiveness, and stability. By surveying existing primitive GFMs and distilling actionable design guidelines for node classification, link prediction, and graph classification, the authors provide a unified framework for future GFM development. They further explore how neural scaling laws may manifest in GFMs, discussing strategies for data augmentation, synthetic graph generation, model scaling, and even the integration of large language models. Altogether, this work offers both a principled perspective and a practical blueprint for advancing GFMs toward general-purpose graph reasoning. This paper inspires many future works, and is selected as a spotlight presentation at ICML 2024. [Text-space Graph Foundation Models: Comprehensive Benchmarks and New Insights](https://arxiv.org/abs/2406.10727) **Zhikai Chen**, Haitao Mao, Jingzhe Liu, Yu Song, Bingheng Li, Wei Jin, Bahare Fatemi, Anton Tsitsulin, Bryan Perozzi, Hui Liu, Jiliang Tang [Code](https://github.com/CurryTang/TSGFM); NeurIPS 2024 Datasets and Benchmarks Track (poster) > [!faq]- Description ### Theoretical Understanding [Unveiling Mode Connectivity in Graph Neural Networks](https://arxiv.org/abs/2502.12608) Bingheng Li, **Zhikai Chen**, Haoyu Han, Shenglai Zeng, Jingzhe Liu, Jiliang Tang; KDD 2025 > [!faq]- Description > ## Contact: [email protected] [Twitter](https://x.com/MathW26496) [Linkedin](https://www.linkedin.com/in/zhikai-chen-435252129/) [Github](https://github.com/CurryTang)