ICLR 2022, 2020
TL;DR: Shows how one party can recover private labels during split training, and proposes Marvell, a noise perturbation that minimizes worst-case leakage.
Jiankai Sun
Machine Learning Engineer at Pinterest · Ph.D. in Computer Science (The Ohio State University)
I am a Machine Learning Engineer at Pinterest. Previously, I was a Research Scientist at ByteDance AML-US in Seattle. I obtained my Ph.D. in Computer Science from The Ohio State University, advised by Prof. Srinivasan Parthasarathy.
My research and engineering focus centers on Privacy-Preserving Machine Learning (Federated Learning, Split Learning, Data Clean Rooms), Large-Scale Recommender Systems (Ads ranking, decoupled representation learning, retrieval), Graph Representation Learning, and Large Language Models & Responsible AI (code generation, proxy fairness). Outside of work, I love vibe coding indie mobile apps, traditional board games, and local-first utilities.
1,639
Citations
20
h-index
23
i10-index
31
Publications
Core Research Themes
Scientific and algorithmic focus areas with corresponding publications.
01 / PRIVACY
12 papers
Federated & Split Learning
Two-party split learning label protection, cross-silo differential privacy, data clean rooms, and vertical federated learning.
View papers →
02 / RECSYS
7 papers
Recommender Systems
Decoupled representation learning for ads ranking, end-to-end retrieval, machine unlearning, and diversified item ranking.
View papers →
03 / GRAPHS
8 papers
Graph Representation
Directed graph embedding with asymmetric transitivity preservation, cycle breaking in noisy hierarchies, and cold-start routing.
View papers →
04 / LLMS
2 papers
LLMs & Responsible AI
Improving LLM code generation through print debugging, proxy fairness with missing sensitive attributes, and self-refinement.
View papers →
Vibe Coding & Indie Software
Handcrafted native iOS apps, board game engines, and local-first personal utilities.
FolkBoardGames
Game Suite
Traditional Folk Board Games
TinyTasteBook
iOS App
Baby Solid Foods Tracker
HoiBaby
iOS Utility
Newborn Feeding Tracker
Publications Stream
31 papers
WSDM 2023
★ HIGHLIGHT
Dgrec: Graph neural network for recommendation with diversified embedding generation
217 citesProceedings of the sixteenth ACM international conference on web search and …, 2023
TL;DR: Diversifies GNN recommenders through submodular neighbor selection, layer attention, and loss reweighting for long-tail categories.
Proceedings of the 26th international conference on World Wide Web, 183-192, 2017
Proceedings of the 2018 SIAM international conference on data mining, 153-161, 2018
CIKM 2021
★ HIGHLIGHT
Learning an end-to-end structure for retrieval in large-scale recommendations
75 citesProceedings of the 30th ACM international conference on information …, 2021
AAAI 2019
★ HIGHLIGHT
ATP: Directed Graph Embedding with Asymmetric Transitivity Preservation
69 citesThe Thirty-Third AAAI Conference on Artificial Intelligence (AAAI 2019), 2018
TL;DR: Embeds directed graphs while preserving asymmetric transitivity, using graph hierarchy and reachability to route questions in CQA sites.
ACM TIST 2014
★ HIGHLIGHT
Vsrank: A novel framework for ranking-based collaborative filtering
66 citesACM Transactions on Intelligent Systems and Technology (TIST) 5 (3), 1-24, 2014
Proceedings of the 2017 ACM on web science conference, 151-160, 2017
TL;DR: Removes cycles from noisy taxonomy graphs by inferring hierarchy with TrueSkill and social agony, preserving structure while deleting few edges.
arXiv preprint arXiv:2203.02073, 2022
2022
★ HIGHLIGHT
Label leakage and protection from forward embedding in vertical federated learning
62 citesarXiv preprint arXiv:2203.01451, 2022
arXiv preprint arXiv:2106.05508, 2021
ICML 2023
★ HIGHLIGHT
Weak proxies are sufficient and preferable for fairness with missing sensitive attributes
44 citesInternational Conference on Machine Learning, 43258-43288, 2023
TL;DR: Shows off-the-shelf proxies can mislead fairness measurement, and gives an algorithm that measures fairness provably accurately using three weak proxies.
arXiv 2024
★ HIGHLIGHT
Leveraging print debugging to improve code generation in large language models
40 citesarXiv preprint arXiv:2401.05319, 2024
TL;DR: Guides LLMs to debug generated code with print statements and log analysis, outperforming rubber-duck debugging on easy and medium LeetCode problems with GPT-4.
FL-ICML 2021
★ HIGHLIGHT
Defending against Reconstruction Attack in Vertical Federated Learning
40 citesFL-ICML 2021, 2021
Proceedings of the 21st ACM international conference on Information and …, 2012
2018
QDEE: Question Difficulty and Expertise Estimation in Community Question Answering Sites
35 citesThe International AAAI Conference on Web and Social Media (ICWSM 2018), 2018
Proceedings of the 21st ACM international conference on Information and …, 2012
2020
EndCold: An End-to-End Framework for Cold Question Routing in Community Question Answering Services
28 citesThe Twenty-Ninth International Joint Conferences on Artificial Intelligence …, 2020
NeurIPS W 2022
★ HIGHLIGHT
Netflix and forget: Efficient and exact machine unlearning from bi-linear recommendations
26 citesarXiv preprint arXiv:2302.06676, 2023
TL;DR: Exact machine unlearning for bi-linear recommenders: Unlearn-ALS removes user data without retraining or model degradation.
ECML PKDD 2018
ColdRoute: effective routing of cold questions in stack exchange sites: J. Sun et al.
26 citesData mining and knowledge discovery 32 (5), 1339-1367, 2018
arXiv preprint arXiv:2301.07284, 2023
The Thirty-Seventh AAAI Conference on Artificial Intelligence (AAAI 2023 …, 2022
TL;DR: An evaluation algorithm that accurately computes AUC under label differential privacy in federated learning.
Proceedings of the 32nd ACM International Conference on Information and …, 2023
TL;DR: Shares only gradients computed on differentially private flipped labels, with an extra sub-model trained on true labels to preserve prediction accuracy.
arXiv preprint arXiv:2205.12412, 2022
Proceedings of the 18th ACM Conference on Recommender Systems, 819-822, 2024
TL;DR: Trains CVR models in data clean rooms using batch-level aggregated gradients, compressed adapter-based fine-tuning, and de-biasing under label differential privacy.
Proceedings of the 29th ACM SIGKDD conference on knowledge discovery and …, 2023
TL;DR: Estimates interaction importance from user-item graph topology via graph conductance and propagation, with no pre-trained pilot model required.
arXiv preprint arXiv:2206.07998, 2022
NeurIPS ML Safety Workshop, 2022
The Thirty-Third AAAI Conference on Artificial Intelligence (AAAI 2019), 2018
Proceedings of the Nineteenth ACM Conference on Recommender Systems, 940-944, 2025
TL;DR: Builds user and Pin embeddings in upstream models, refreshed asynchronously and reused as features across downstream ad retrieval and CTR/CVR ranking in production at Pinterest.
The Ohio State University, 2019