Zexi Huang

• Course GPA: 4.0/4.0. UCSB Computer Science Outstanding Scholar Fellow (top 4 out 63 PhD students).
• Advisor: Prof. Ambuj Singh. Dissertation: Learning Representations for Information-rich Graphs.

• GPA: 3.96/4.0, Avg. Score: 92.79/100 (1/90 in freshman year, 1/87 in sophomore year, 1/93 in junior year).
• Advisors: Prof. Junming Shao and Prof. Sinno Jialin Pan. Thesis: Transfer Learning for Community Detection in Multiplex Networks.

• GPA:4.0/4.0, Avg. Score: 96.44/100.

• Leading a cross-functional team of machine learning, infrastructure, and platform engineers to design, build, and scale next-generation recommender systems, owning key initiatives including VLM/LRM co-training, LLM4Rec, and generative recommendation.

• Designed and deployed a novel feature interaction architecture that scales up the ranking baseline by 50x, achieving +1%/-1% in CTR AUC/Logloss and +10%/+8% in Quality Watch Session/Duration / User, with superior scaling laws compared to state-of-the-art methods.
• Incorporated user behavioral information from other domains (short videos, e-commerce, and search) with graph representation learning.
• Encoded visual, ASR, and metadata information from the content with vision-language models to improve model generalizability.

• Owned the iterations of the core ranking models of TikTok Live Recommendation, leveraging representation learning, multi-task learning, contrastive learning, regression modeling, and sequence modeling, with +7% Watch Duration / User gains from online A/B experiments.
• Developed the host go-live model that captures the relationship between watch-live and go-live with causal inference and uplift modeling to motivate authorized hosts to go-live, achieving +2% Host Session / User in online A/B experiments.

• Developed a stochastic inventory management model based on dynamic programming for Amazon’s Print-On-Demand business, optimizing the ordering strategy for 42.4M units of books and realizing an annual saving of $10.4M.
• Designed and implemented a graph-based NLP model for long text embedding and classification using graph neural networks, leading to superior performance to state-of-the-art transformer-based models for Kindle book contents.

• Designed graph-based machine learning models for fraud detection based on multi-modal signals in Kindle Direct Publishing.
• Implemented the models in an end-to-end fashion and deployed them to graphs with millions of nodes and billions of edges in production.
• Validation results show that the models can surface fraud rings undetected by existing processes with an estimated annual value of $2.4M.

• Proposed to augment existing fraud detection methods with graph-based machine learning models for Kindle Direct Publishing.
• Designed and implemented various heuristics and an embedding framework for attributed heterogeneous multiplex networks.
• The models are deployed into production and results show that they surface up to 15 times more fraud compared to the existing processes.

• Proposed to refine community detection results in some layers with transferred knowledge from other layers in multiplex networks.
• Designed a representation-based community detection framework and implemented it with an extended symmetric NMF approach.
• Our algorithm outperforms other representation-based community detection algorithms, especially when the target layer is noisy.

[Preprint] Ruifeng Zhang*, Zexi Huang*†, Zikai Wang*, Ke Sun* et al. Zenith: Scaling up Ranking Models for Billion-scale Livestreaming Recommendation
. arXiv preprint arXiv:2601.21285, 2026. (*: equal contribution, †: corresponding author)

[TIST-2025] Mert Kosan*, Zexi Huang*, Sourav Medya, Sayan Ranu, Ambuj Singh. GCFExplainer: Global Counterfactual Explainer for Graph Neural Networks. ACM Transactions on Intelligent Systems and Technology, 2025. (*: equal contribution) [Code] [Slides] [Talk]

[AAAI-2024] Aritra Bhowmick, Mert Kosan, Zexi Huang, Sourav Medya, Ambuj Singh, Sourav Medya. DGCLUSTER: A Neural Framework for Attributed Graph Clustering via Modularity Maximization. AAAI Conference on Artificial Intelligence, 2024. [Code]

[WSDM-2023] Zexi Huang*, Mert Kosan*, Sourav Medya, Sayan Ranu, Ambuj Singh. Global Counterfactual Explainer for Graph Neural Networks. ACM International Conference on Web Search and Data Mining, 2023. (*: equal contribution) [Code] [Slides] [Talk]

[WSDM-2022] Zexi Huang, Arlei Silva, Ambuj Singh. POLE: Polarized Embedding for Signed Networks. ACM International Conference on Web Search and Data Mining, 2022. [Code] [Poster] [Slides] [Talk]

[KDD-2021] Zexi Huang, Arlei Silva, Ambuj Singh. A Broader Picture of Random-walk Based Graph Embedding. ACM SIGKDD Conference on Knowledge Discovery & Data Mining, 2021. [Code] [Poster] [Slides] [Talk] [Talk (in Chinese)]

[Preprint] Zexi Huang, Mert Kosan, Arlei Silva, Ambuj Singh. Link Prediction without Graph Neural Networks. arXiv preprint arXiv:2305.13656, 2023. [Code]

[Preprint] Wei Ye, Zexi Huang, Yunqi Hong, Ambuj Singh. Graph Neural Diffusion Networks for Semi-supervised Learning. arXiv preprint arXiv:2201.09698, 2022. [Code]

• Scrutinized the training and evaluation of link prediction methods and identify their limitations in handling class imbalance.
• Proposed a novel topology-centric framework that combines graph learning, topological heuristics, and an N-pair loss for link prediction.
• Results showed that the proposed method is 145% more accurate and trains/infers 11/6,000 times faster than the state-of-the-art methods.

• Formulated the novel problem of global counterfactual reasoning/explanation of graph neural networks for graph classification.
• Proposed GCFExplainer, the first global explainer powered by vertex-reinforced random walks on an edit map with a greedy summary.
• Results showed that GCFExplainer not only provides crucial high-level insights but also outperforms existing methods in recourse quality.

• Designed a novel polarization measure for signed graphs and showed that existing methods fail in polarized signed link prediction.
• Proposed a polarized embedding algorithm that captures both topological and signed similarity jointly via signed autocovariance.
• Extensive experiments showed that the proposed model outperforms state-of-the-art methods by up to one order of magnitude.

• Interpreted the layer-wise propagation rule of GCN from the perspective of power iteration and analyzed its converging process.
• Designed a novel GCN architecture that learns to aggregate multiscale information based on graph diffusions with a neural network.
• Illustrated the effectiveness and efficiency of the proposed model by extensive comparative studies with state-of-the-art methods.

• Presented a unified view of embedding, covering different random-walk processes, similarity metrics, and embedding algorithms.
• Showed both theoretical and empirical evidence of the superiority of the novel autocovariance embedding in link prediction.
• Illustrated ways to exploit the multiscale nature of random-walk similarity to further optimize embedding performance.

• Proposed an intuitive algorithm for fast overlapping community detection in networks.
• Introduced gain and loss functions from game theory to model the intrinsic dynamics between nodes.
• Results are comparable to state-of-the-art algorithms.

• KDD’23

• AAAI’23-25
• KDD’22
• SDM’22

• TNNLS’23
• Neural Networks’23
• TIST’22-26
• TKDD’21-24
• KDD’20-21
• WebConf’21

• Graduate Affairs Committee, Department of Computer Science, UCSB

• SB Hacks V, VII, VIII Hackathon