Yanjie Fu

AI ⇄ Data

AI systems, unlike humans, are brittle, not robust, often struggle when faced with novel situations, and highly sensitive to small perturbations, which can lead to catastrophically poor performance. My research aims to develop robust machine intelligence with imperfect and complex data, by building tools to address learning framework, algorithmic, data, and computing challenges. My research involves two major efforts: 1) on the AI side, I study how the structure knowledge of data can guide AI; 2) on the data side, I examine how AI can transform, reprogram, augment, and knowledgize data. I execute two important steps steps towards this vision. The first step (data representation construct) aims to integrate structure knowledge to achieve deep robust representation to fight imperfect and complexity data. The second step (learning strategy construct) aims to integrate robust representations with adaptive and interactive learning to fight uncertain and constrained environments.

A. AI For The Data: AI for Data Augmentation, Reprogramming, Knowlegization

A-1 Data-centric AI: from Reinforcement Decisions to Generative Intelligence

• Key concepts: learning mechanism unknown feature representation space knowledge via generative modeling and continuous latent reasoning for data editing
• Ph.D. Dissertation Student: Dongjie Wang (2020-2024), a tenure-track assistant professor at the University of Kansas
• Key papers:
  • Group-wise Reinforcement Feature Generation for Optimal and Explainable Representation Space Reconstruction (KDD 2022)
  • Reinforcement-Enhanced Autoregressive Feature Transformation: Gradient-steered Search in Continuous Space for Postfix Expressions (NIPS 2023 Spotlight)

A-2 Reinforcement Learning and Automated Data Science

• Key concepts: self optimizing feature selection and generation as decision sequences of reinforcement policy networks
• Ph.D. Dissertation Student: Kunpeng Liu (2017-2022), a tenure-track assistant professor at Clemson University
• Key papers:
  • Automated Feature Selection: A Reinforcement Learning Perspective (TKDE)
  • Interactive Reinforcement Learning for Feature selection with Decision Tree in the Loop. (TKDE)
  • Efficient Reinforced Feature Selection via Early Stopping Traverse Strategy. (ICDM)
  • Automating Feature Subspace Exploration via Multi-Agent Reinforcement Learning. (KDD)
• Awards
  • IEEE ICDM Best Paper Finalist

B. AI By The Data: Data’s Structure Knowledge for Guiding AI

B-1 Deep Time Series Learning

• Key concepts: from time series regularities to time series shifts
• Ph.D. Dissertation Student: Wei Fan (2020-2023), a tenure-track assistant professor at University of Auckland
• Key papers:
  • DEPTS: Deep Expansion Learning for Periodic Time Series Forecasting. (ICLR22)
  • Dish-TS: A General Paradigm for Alleviating Distribution Shift in Time Series Forecasting. (AAAI23)
  • IN-Flow: Address Distribution Shift in Time Series Forecasting. (KDD2025)

B-2 Structure Knowledge Guided Spatial-Temporal Representation Learning

• Key concepts: integrating spatial-temporal knowledge guidance (e.g., collective, peer, dynamic, substructure knowledge) with graph embedding to amplify spatiotemporal representation learning
• Ph.D. Dissertation Student: Pengyang Wang (2017-2021), a tenure-track assistant professor at University of Macau
• Key papers:
  • Mutual Information Based Substructured Representation Learning (IJCAI20)
  • Adversarial Substructured Representation Learning (KDD19)
  • Peer and Temporal-Aware Representation Learning (KDD’18)
• Awards:
  • ACM SIGKDD’18 Best Student Paper Finalist

B-3 Time Series and Root Cause Analysis

• Key concepts: learning causal structures from multivariate time series
• Ph.D. Students: Dr. Dongjie Wang, Arun Vignesh
• Key papers:
  • Incremental causal graph learning for online root cause analysis (KDD 2023)
  • Interdependent causal networks for root cause localization (KDD 2023)

C. AI Beyond The Data: AI to Reason Beyond Data Distribution and Dynamics as Extrapolation Knowledge for Simulation, Decisions, World Models

• Key concepts: simulation as generative AI; decision making as policy learning or agentic reasoning; building spatial, time, social, and physical knowledge-guided digital copy of complex system distribution and dynamics; reason decision intelligence from multimodal semi-structure hybrid data
• Ph.D. Students: Haoyue Bai
• Key papers:
  • Generative Simulation and Iterative Decision Policies for Supply Chain Optimization (INFORMS Winter Simulation Conference 2025)
  • Brownian Bridge Augmented Surrogate Simulation and Injection Planning for Geological CO2 Storage (NeurIPS 2025 underreview)

AI Interactions: Embedding AI into Critical Systems and Interdisciplinary Research

• AI and Urban Planning
  • Key concepts: urban planning as land-use configuration generation
  • Ph.D. Students: Dr. Dongjie Wang
  • Community Building: AAAI 25 urban planning workshop, 25 AI and Cities Forum
  • Key papers:
    • Towards Automated Urban Planning: When Generative and ChatGPT-like AI Meets Urban Planning
    • Reimagining city configuration: Automated urban planning via adversarial learning (SIGSPatial 20)
    • Human-instructed Deep Hierarchical Generative Learning for Automated Urban Planning (AAAI 23)
    • Automated Urban Planning for Reimagining City Configuration via Adversarial Learning: Quantification, Generation, and Evaluation (ACM Trans on SAS 23)
    • Deep human-guided conditional variational generative modeling for automated urban planning (IEEE ICDM 23)
    • Hierarchical Reinforced Urban Planning: Jointly Steering Region and Block Configurations (SIAM DM 23)
  • Awards:
    • ACM SIGSpatial best paper runner-up
• Human Mobility Modeling
  • Key concepts: human mobility as stochastic processes, word/topic modeling
  • Human Mobility Synchronization and Trip Purpose Detection with Mixture of Hawkes Processes (KDD’17)
  • Representing Urban Forms: A Collective Learning Model with Heterogeneous Human Mobility Data (TKDE)
  • Human Mobility Synchronization and Trip Purpose Detection with Mixture of Hawkes Processes (KDD’17)
• Machine Learning for Urban Vibrancy Analysis with Crowd-sourced Geo-tagged Data
  • Mixed Land-use Latent Models for Geographic Ranking (KDD’15)
  • Exploiting Geographic Dependencies for Residential Community Ranking (KDD’14)
• Machine Learning for In-App Behavior Analysis
  • A Multi-Label Multi-View Learning Framework for In-App Service Usage Analysis (TIST)
  • Effective and Real-time In-App Activity Analysis in Encrypted Internet Traffic Streams (KDD’17)
  • Service Usage Classification with Encrypted Internet Traffic in Mobile Messaging Apps (TMC)
  • Service Usage Analysis in Mobile Messaging Apps: A Multi-Label Multi-View Perspective (ICDM’16)
• Machine Learning for Mobile Recommender Systems
  • Unsupervised P2P Rental Recommendations via Integer Programming (KDD’17)
  • User Preference Learning With Multiple Information Fusion For Restaurant Recommendation (SDM’14)
  • Learning Geographical Preferences for Point-of-Interest Recommendation (KDD’13)