We propose a novel stochastic process that is with probability $alpha_i$ being absorbed at current state $i$, and with probability $1-alpha_i$ follows a random edge out of it. We analyze its properties and show its potential for exploring graph structures. We prove that under proper absorption rates, a random walk starting from a set $mathcal{S}$ of low conductance will be mostly absorbed in $mathcal{S}$. Moreover, the absorption probabilities vary slowly inside $mathcal{S}$, while dropping sharply outside, thus implementing the desirable cluster assumption for graph-based learning. Remarkably, the partially absorbing process unifies many popular models arising in a variety of contexts, provides new insights into them, and makes it possible for transferring findings from one paradigm to another. Simulation results demonstrate its promising applications in retrieval and classification.

We develop a hybrid social media network to integrate the diverse types of entities and relations in an optimal selective way for each query adaptively. Rather than weight averaging multiple edges or treating diverse relation types as the same in the inferencing process, we propose an optimal relation selection method based on several sparsity principles. The model not only allows effective information diffusion but also answer the question - which node and relation play the most important role in diffusing information for each query and task. Results of a few applications (image tagging, target multimedia ad, text-to-image illustration) over the real social multimedia data sets confirm the clear performance gains.

In this paper, we propose a brand new application of hashing techniques - accelerating the common bottleneck operation, min/max inner product, encountered in many large-scale optimization algorithms used in applications like active learning SVM, sparse coding, and other computer vision tasks. Our technique is based on some unique properties of order statistics computed from statistically correlated random vectors.

We develop a new visual representation, called scene aligned pooling, for complex video event recognition. The key idea is to align each frame in a video with component scenes, which tend to capture unique characteristics of an event and can be discovered automatically. Feature pooling and event classification models are then performed over the aligned scenes. Results of TRECVID MED and the Human Motion Recognition Databases (HMDB) demonstrate the significant peroformance gains. 

Graph-based semi-supervised learning (GSSL) provides a promising paradigm for modeling the manifold structures that may exist in massive data sources in high-dimensional spaces. This paper provides a survey of several classical GSSL methods and a few promising methods in handling challenging issues often encountered in web-scale applications, including how to handle noisy labels and gigantic data sets.

We developed a novel method for retrieving images over a large pool of weak attributes (> 6000), which can be imported from external sources and do not require manual annotations. We used a semi-supervised learning approach to map semantic attribute-based queries to weak attributes and determine the optimal retrieval weights.

We present a low-rank reconstruction method to address the domain distribution disparity when adapting image classification models across domains. Besides clear performance improvement, our method has the unique capability in uncovering outlier points that cannot be adapted.

We propose a rank minimization method to fuse the predicted confidence scores of multiple models, each of which is obtained based on a certain kind of feature. 

We proposed a novel bipartite formulation to integrate superpixels extracted by different models at different scales, and a highly efficient linear-time spectral solver to demonstrate significant better image segmentation results.

We developed a new multiple source domain adaptation method called Domain Selection Machine (DSM) for event recognition in consumer videos by leveraging a large number of loosely labeled web images from different sources such as Flickr. A performance gain of 46% is observed.

In this work, we propose a novel mobile visual search system based on ”Bag of Hash Bits” (BoHB), in which each local feature is encoded to a very small number of hash bits, instead of quantized to visual words, and the whole image is represented as bag of hash bits.  We also incorporate a boundary feature in the reranking step to describe the object shapes, complementing the local features that are usually used to characterize the local details.

In this paper we propose a novel hashing function based on hypershpere, Spherical Hashing, to map more spatially coherent data points into the same binary code compared to hyperplane based hashing functions.The excellent results confirm the unique merits of the proposed idea in using hyperspheres to encode proximity regions in high dimensional spaces.

We showed the supervised information such as semantic label consistence and metric neighborhood between pairs of points can be utilized to design robust compact hash codes. Our supervised hashing technique is unique in using code inner product rather than Hamming distance and incorporating kernel-based models in the optimization process.

We develop a joint audio-visual bi-modal representation  to discover strong audio-visual joint patterns  in videos for detecting multimedia events.

This paper addresses a new problem of query-adaptive hashing. Our work is distinct from others by two unique features: 1) mixed image-keyword query, namely each query is comprised of an examplar image and several descriptive keywords, and 2) each data point is associated with multiple labels. We propose a boosting style algorithm for effectively learning compact hash codes under above settings.

Hyperplane hashing aims at rapidly searching nearest points to a hyperplane, and has shown major impact in scaling up active learning with SVMs. We developed a projection technique based on a novel bilinear form, with a proven higher collision probability than all prior results. We further developed methods to learn the bilinear functions directly from the data.

How difficult is (approximate) nearest neighbor search in a given data set? And which data properties affect the difficulty of nearest neighbor search and how?This paper introduces the first concrete measure called Relative Contrast that can be used to evaluate the influence of several crucial data characteristics such as dimensionality, sparsity, and database size simultaneously in arbitrary normed metric spaces.