Dongjin Song

Biologically Inspired Feature Manifold for Scene Classification

Biologically inspired feature (BIF) and its variations have been demonstrated to be effective and efficient for scene classification. It is unreasonable to measure the dissimilarity between two BIFs based on their Euclidean distance. This is because BIFs are extrinsically very high dimensional and intrinsically low dimensional, i.e., BIFs are sampled from a low-dimensional manifold and embedded in a high-dimensional space. Therefore, it is essential to find the intrinsic structure of a set of BIFs, obtain a suitable mapping to implement the dimensionality reduction, and measure the dissimilarity between two BIFs in the low-dimensional space based on their Euclidean distance. In this paper, we study the manifold constructed by a set of BIFs utilized for scene classification, form a new dimensionality reduction algorithm by preserving both the geometry of intra BIFs and the discriminative information inter BIFs termed Discriminative and Geometry Preserving Projections (DGPP), and construct a new framework for scene classification. In this framework, we represent an image based on a new BIF, which combines the intensity channel, the color channel, and the C1 unit of a color image; then we project the high-dimensional BIF to a low-dimensional space based on DGPP; and, finally, we conduct the classification based on the multiclass support vector machine (SVM). Thorough empirical studies based on the USC scene dataset demonstrate that the proposed framework improves the classification rates around 100% relatively and the training speed 60 times for different sites in comparing with previous gist proposed by Siagian and Itti in 2007.

High resolution population estimates from telecommunications data

Spatial variations in the distribution and composition of populations inform urban development, health-risk analyses, disaster relief, and more. Despite the broad relevance and importance of such data, acquiring local census estimates in a timely and accurate manner is challenging because population counts can change rapidly, are often politically charged, and suffer from logistical and administrative challenges. These limitations necessitate the development of alternative or complementary approaches to population mapping. In this paper we develop an explicit connection between telecommunications data and the underlying population distribution of Milan, Italy. We go on to test the scale invariance of this connection and use telecommunications data in conjunction with high-resolution census data to create easily updated and potentially real time population estimates in time and space.

Top Rank Supervised Binary Coding for Visual Search

In recent years, binary coding techniques are becoming increasingly popular because of their high efficiency in handling large-scale computer vision applications. It has been demonstrated that supervised binary coding techniques that leverage supervised information can significantly enhance the coding quality, and hence greatly benefit visual search tasks. Typically, a modern binary coding method seeks to learn a group of coding functions which compress data samples into binary codes. However, few methods pursued the coding functions such that the precision at the top of a ranking list according to Hamming distances of the generated binary codes is optimized. In this paper, we propose a novel supervised binary coding approach, namely Top Rank Supervised Binary Coding (Top-RSBC), which explicitly focuses on optimizing the precision of top positions in a Hamming-distance ranking list towards preserving the supervision information. The core idea is to train the disciplined coding functions, by which the mistakes at the top of a Hamming-distance ranking list are penalized more than those at the bottom. To solve such coding functions, we relax the original discrete optimization objective with a continuous surrogate, and derive a stochastic gradient descent to optimize the surrogate objective. To further reduce the training time cost, we also design an online learning algorithm to optimize the surrogate objective more efficiently. Empirical studies based upon three benchmark image datasets demonstrate that the proposed binary coding approach achieves superior image search accuracy over the state-of-the-arts.

Efficient Latent Link Recommendation in Signed Networks

Signed networks, in which the relationship between two nodes can be either positive (indicating a relationship such as trust) or negative (indicating a relationship such as distrust), are becoming increasingly common. A plausible model for user behavior analytics in signed networks can be based upon the assumption that more extreme positive and negative relationships are explored and exploited before less extreme ones. Such a model implies that a personalized ranking list of latent links should place positive links on the top, negative links at the bottom, and unknown status links in between. Traditional ranking metrics, e.g., area under the receiver operating characteristic curve (AUC), are however not suitable for quantifying such a ranking list which includes positive, negative, and unknown status links. To address this issue, a generalized AUC (GAUC) which can measure both the head and tail of a ranking list has been introduced. Since GAUC weights each pairwise comparison equally and the calculation of GAUC requires quadratic time, we derive two lower bounds of GAUC which can be computed in linear time and put more emphasis on ranking positive links on the top and negative links at the bottom of a ranking list. Next, we develop two efficient latent link recommendation (ELLR) algorithms in order to recommend links by directly optimizing these two lower bounds, respectively. Finally, we compare these two ELLR algorithms with top-performing baseline methods over four benchmark datasets, among which the largest network has more than 100 thousand nodes and seven million entries. Thorough empirical studies demonstrate that the proposed ELLR algorithms outperform state-of-the-art approaches for link recommendation in signed networks at no cost in efficiency.

C1 units for scene classification

In this paper, we unify C1 units and the locality preserving projections (LPP) into the conventional gist model for scene classification. For the improved gist model, we first utilize the C1 units, intensity channel and color channel of color image to represent the color image with the high dimensional feature, then we project high dimensional samples to a low dimensional subspace via LPP to preserve both the local geometry and the discriminate information, and finally, we apply the nearest neighbour rule with the Euclidean distance for classification. Experimental results based on the USC scene database not only demonstrate that the proposed gist improves the classification accuracy around 7% but also reduce the testing cost around 50 times in comparing with the original gist model proposed by Siagian and Itti in TPAMI 2007.

Rank Preserving Hashing for Rapid Image Search

In recent years, hashing techniques are becoming overwhelmingly popular for their high efficiency in handling large-scale computer vision applications. It has been shown that hashing techniques which leverage supervised information can significantly enhance performance, and thus greatly benefit visual search tasks. Typically, a modern hashing method uses a set of hash functions to compress data samples into compact binary codes. However, few methods have developed hash functions to optimize the precision at the top of a ranking list based upon Hamming distances. In this paper, we propose a novel supervised hashing approach, namely Rank Preserving Hashing (RPH), to explicitly optimize the precision of Hamming distance ranking towards preserving the supervised rank information. The core idea is to train disciplined hash functions in which the mistakes at the top of a Hamming-distance ranking list are penalized more than those at the bottom. To find such hash functions, we relax the original discrete optimization objective to a continuous surrogate, and then design an online learning algorithm to efficiently optimize the surrogate objective. Empirical studies based upon two benchmark image datasets demonstrate that the proposed hashing approach achieves superior image search accuracy over the state-of-the-art approaches.

Link sign prediction and ranking in signed directed social networks

Signed directed social networks, in which the relationships between users can be either positive (indicating relations such as trust) or negative (indicating relations such as distrust), are increasingly common. Thus the interplay between positive and negative relationships in such networks has become an important research topic. Most recent investigations focus upon edge sign inference using structural balance theory or social status theory. Neither of these two theories, however, can explain an observed edge sign well when the two nodes connected by this edge do not share a common neighbor (e.g., common friend). In this paper, we develop a novel approach to handle this situation by applying a new model for node types and use the proposed model to perform link sign prediction and link ranking. Initially, we analyze the local node structure in a fully observed signed directed network, inferring underlying node types. The sign of an edge between two nodes must be consistent with their types; this explains edge signs well even when there are no common neighbors. We show, moreover, that our approach can be extended to incorporate directed triads, when they exist, just as in models based upon structural balance or social status theory. We compute Bayesian node types within empirical studies based upon partially observed Wikipedia, Slashdot, and Epinions networks in which the largest network (Epinions) has 119K nodes and 841K edges. Based upon the proposed features, we present the link sign prediction and link ranking models subsequently. We show that our approaches yield better performance than state-of-the-art approaches for these two tasks based upon three signed directed networks.

A model of consistent node types in signed directed social networks

Signed directed social networks, in which the relationships between users can be either positive (indicating relations such as trust) or negative (indicating relations such as distrust), are increasingly common. Thus the interplay between positive and negative relationships in such networks has become an important research topic. Most recent investigations focus upon edge sign inference using structural balance theory or social status theory. Neither of these two theories, however, can explain an observed edge sign well when the two nodes connected by this edge do not share a common neighbor (e.g., common friend). In this paper we develop a novel approach to handle this situation by applying a new model for node types. Initially, we analyze the local node structure in a fully observed signed directed network, inferring underlying node types. The sign of an edge between two nodes must be consistent with their types; this explains edge signs well even when there are no common neighbors. We show, moreover, that our approach can be extended to incorporate directed triads, when they exist, just as in models based upon structural balance or social status theory. We compute Bayesian node types within empirical studies based upon partially observed Wikipedia, Slashdot, and Epinions networks in which the largest network (Epinions) has 119K nodes and 841K edges. Our approach yields better performance than state-of-the-art approaches for these three signed directed networks.

Efficient Robust Conditional Random Fields

Conditional random fields (CRFs) are a flexible yet powerful probabilistic approach and have shown advantages for popular applications in various areas, including text analysis, bioinformatics, and computer vision. Traditional CRF models, however, are incapable of selecting relevant features as well as suppressing noise from noisy original features. Moreover, conventional optimization methods often converge slowly in solving the training procedure of CRFs, and will degrade significantly for tasks with a large number of samples and features. In this paper, we propose robust CRFs (RCRFs) to simultaneously select relevant features. An optimal gradient method (OGM) is further designed to train RCRFs efficiently. Specifically, the proposed RCRFs employ the

Discrminative Geometry Preserving Projections

\ell _{1}