Graph-based semi-supervised learning (GSSL) provides
a promising paradigm for modeling the manifold structures that often exist in massive
data in high-dimensional spaces. It has been shown effective in propagating a limited
amount of initial labels to a large amount of unlabeled data, matching the needs
of many emerging applications such as image annotation and information retrieval.
We have developed a family of techniques to solve the open problems such as unbalanced
labels, contaminated noisy labels, and graph construction over gigantic datasets.
We have applied such techniques to many real-world applications such as interactive
image retrieval, noisy Web image reranking, bio-molecular cellular image mining,
and brain machine interfaces for image retrieval. We have also combined the graph-based
learning and hashing techniques to derive graph-based hashing codes for scalable
- Jun Wang, Tony Jebara, Shih-Fu Chang. Graph Transduction via Alternating Minimization.
In International Conference on Machine Learning (ICML), Helsinki, Finland, July
We proposed a bi-variate alternate optimization technique to derive the optimal
prediction function over graphs. It treated both the initial labels and predicted
function as optimization variables. It was shown effective in handling unbalanced
and noisy label conditions.
- Tony Jebara, Jun Wang, Shih-Fu Chang. Graph Construction and b-Matching for Semi-Supervised
Learning. In International Conference on Machine Learning (ICML), Montreal, Canada,
June 2009. [pdf]
This paper shows graph construction using b-matching is better than those constructed
- Jun Wang, Yu-Gang Jiang, Shih-Fu Chang. Label Diagnosis through Self Tuning for
Web Image Search. In IEEE Computer Society Conference on Computer Vision and Pattern
Recognition (CVPR), Miami Beach, Florida, USA, June 2009. [pdf]
We demonstrated promising results in filtering and denoising the incorrectly labeled
images from the Web by graph-based diagnosis and tuning.
- Wei Liu, Junfeng He, Shih-Fu Chang. Large Graph Construction for Scalable Semi-Supervised
Learning. In the 27th International Conference on Machine Learning (ICML), Haifa,
Israel, June 2010. [pdf][code]
We proposed a highly efficient method using Anchor Graph for constructing sparse
low-rank graphs and semi-supervised learning with only linear complexity over gigantic
- Wei Liu, Jun Wang, Sanjiv Kumar, Shih-Fu Chang. Hashing with Graphs. In International
Conference on Machine Learning (ICML), Bellevue, WA, USA, 2011. [pdf]
We applied the Anchor Graph method to derive the eigenfunction over large graphs
without needing over-simplified assumptions about data distributions. We designed
graph-based hashing for large-scale similarity retrieval and demonstrated a retrieval
accuracy even better than that of L2 linear scan.
- Jun Wang, Shih-Fu Chang, Xiabo Zhou, T. C. Stephen Wong. Active Microscopic Cellular
Image Annotation by Superposable Graph Transduction with Imbalanced Labels. In IEEE
Computer Society Conference on Computer Vision and Pattern Recognition (CVPR), Anchorage,
Alaska, USA, June 2008. [pdf]
In this paper, we developed a real-time system for interactive cellular image annotation
using the graph-based SSL method. We specifically applied graph superposition and
normalization ideas to achieve the real-time speed and solve the class imbalance
- Yu-Gang Jiang, Jun Wang, Shih-Fu Chang, Chong-Wah Ngo. Domain Adaptive Semantic
Diffusion for Large Scale Context-Based Video Annotation. In International Conference
on Computer Vision (ICCV), Kyoto, Japan, September 2009. [pdf]
We applied graph-based diffusion to fuse results of individual concept detectors
to improve the overall accuracy of image annotation.
- Jun Wang, Eric Pohlmeyer, Barbara Hanna, Yu-Gang Jiang, Paul Sajda, Shih-Fu Chang.
Brain State Decoding for Rapid Image Retrieval. In Proceeding of the ACM international
conference on Multimedia (ACM MM), October 2009. [pdf]
We combine the EEG-based brain signal decoder and graph-based semi-supervised learning
to detect arbitrary user-initiated search targets and retrieve relevant images from
a large database using a brain machine interface.