P-GNN: Position-aware Graph Neural Networks
Hi! I am a PhD student in Department of Computer Science, Stanford University. I'm advised by Prof. Jure Leskovec.
In the past, I have been fortunate enough to work with Prof. Stefano Ermon and Prof. David Lobell as a summer intern. In Tsinghua, I have worked on Machine Learning supervised by Prof. Jun Zhu.
My research focuses on developing Machine Learning algorithms for graph/relational structured data, including generating/optimizing graph structures, developing new graph neural network architectures, and applying these techniques in various domains.
Existing dynamic recommender systems often use out-of-the-box sequence models which are limited by speed and memory consumption, are often infeasible for production environments, and usually do not incorporate cross-session information. Here we propose Hierarchical Temporal Convolutional Networks (HierTCN), a hierarchical deep learning architecture that makes dynamic recommendations based on users' sequential multi-session interactions with items.
We conduct extensive experiments on a public XING dataset and a large-scale Pinterest dataset of 6 million users with 1.6 billion interactions. We show that HierTCN is 2.5x faster than RNN-based models and uses 90% less data memory compared to TCN-based models.
Generating novel graph structures that optimize given objectives while obeying
some given underlying rules is fundamental for chemistry, biology and social
science research. Here we propose Graph Convolutional Policy
Network (GCPN), a general graph convolutional network based model for goaldirected
graph generation through reinforcement learning. The model is trained
to optimize domain-specific rewards and adversarial loss through policy gradient,
and acts in an environment that incorporates domain-specific rules. Experimental
results show that GCPN can achieve 61% improvement on chemical property
optimization over state-of-the-art baselines while resembling known molecules,
and achieve 184% improvement on the constrained property optimization task.
Recently, graph neural networks (GNNs) have revolutionized the field of graph representation learning through effectively learned node embeddings. Here we propose DiffPool, a differentiable graph pooling module that can generate hierarchical representations of graphs and can be combined with various graph neural network architectures in an end-to-end fashion. Our experimental results show that combining existing GNN methods with DiffPool yields an average improvement of 5-10% accuracy on graph classification benchmarks, compared to all existing pooling approaches, achieving a new state-of-the-art on four out of five benchmark datasets.
Modeling and generating graphs is fundamental
for studying networks in biology, engineering,
and social sciences. Here we
propose GraphRNN, a deep autoregressive model
that addresses the above challenges and approximates
any distribution of graphs with minimal
assumptions about their structure.
In order to quantitatively evaluate
the performance of GraphRNN, we introduce
a benchmark suite of datasets, baselines and novel
evaluation metrics based on Maximum Mean Discrepancy,
which measure distances between sets
of graphs. Our experiments show that GraphRNN
significantly outperforms all baselines, learning to
generate diverse graphs that match the structural
characteristics of a target set, while also scaling
to graphs 50× larger than previous deep models.
Crop yield prediction is central in ensuring the food security. We introduce the first deep learning based method to predict crop yield purely based on publicly available remote sensing data. A novel dimensionality reduction method is proposed based on histogram calculation. By combining Gaussian Process with Convolutional Neural Networks, the model significantly outperforms traditional models, and we show that the features automatically extracted by deep learning models are much more effective than traditional hand-crafted features.
[PDF] [Code] [Project Webpage]
Metro systems are playing an increasing role in
modern cities, and various management solutions are proposed
to improve their efficiency. When analyzing the effects of those
solutions, simulation proves to be an effective and cost-saving
method. This paper presents an effective simulation model for
multi-line metro systems based on the OD (origin-destination)
data and the network connection data. The model is validated in
the scenario of Beijing metro system, which proves its
effectiveness in large-scale empirical implementation.