Matrix factorization based techniques, such as nonnegative matrix factorization (NMF) and concept factorization (CF), have attracted great attention in dimension reduction and data clustering. Both of them are linear learning problems and lead to a sparse representation of the data. However, the sparsity obtained by these methods does not always satisfy locality conditions, thus the obtained data representation is not the best. This paper introduces a locality-constrained concept factorization method which imposes a locality constraint onto the traditional concept factorization. By requiring the concepts (basis vectors) to be as close to the original data points as possible, each data can be represented by a linear combination of only a few basis concepts. Thus our method is able to achieve sparsity and locality at the same time. We demonstrate the effectiveness of this novel algorithm through a set of evaluations on real world applications.

Dimensionality reduction is a very important topic in machine learning. It can be generally classified into two categories: feature selection and subspace learning. In the past decades, many methods have been proposed for dimensionality reduction. However, most of these works study feature selection and subspace learning independently. In this paper, we present a framework for joint feature selection and subspace learning. We reformulate the subspace learning problem and use L {2,1} -norm on the projection matrix to achieve row-sparsity, which leads to selecting relevant features and learning transformation simultaneously. We discuss two situations of the proposed framework, and present their optimization algorithms. Experiments on benchmark face recognition data sets illustrate that the proposed framework outperforms the state of the art methods overwhelmingly.

In this paper, we tackle the challenges of multi-label classification by developing a general conditional dependency network model. The proposed model is a cyclic directed graphical model, which provides an intuitive representation for the dependencies among multiple label variables, and a well integrated framework for efficient model training using binary classifiers and label predictions using Gibbs sampling inference. Our experiments show the proposed conditional model can effectively exploit the label dependency to improve multi-label classification performance.

The RoboCup robot soccer Small Size League has been running since 1997 with many teams successfully competiting and very effectively playing the games. Teams of five robots, with a combined autonomous centralized perception and control, and distributed actuation, move at high speeds in the field space, actuating a golf ball by passing and shooting it to aim at scoring goals. Most teams run their own pre-defined team strategies, unknown to the other teams, with flexible game-state dependent assignment of robot roles and positioning. However, in this fast-paced noisy real robot league, recognizing the opponent team strategies and accordingly adapting one's own play has proven to be a considerable challenge. In this work, we analyze logged data of real games gathered by the CMDragons team, and contribute several results in learning and responding to opponent strategies. We define episodes as segments of interest in the logged data, and introduce a representation that captures the spatial and temporal data of the multi-robot system as instances of geometrical trajectory curves. We then learn a model of the team strategies through a variant of agglomerative hierarchical clustering. Using the learned cluster model, we are able to classify a team behavior incrementally as it occurs. Finally, we define an algorithm that autonomously generates counter tactics, in a simulation based on the real logs, showing that it can recognize and respond to opponent strategies.

This paper describes a novel framework to jointly learn data-dependent label and locality-preserving projections. Given a set of data instances from multiple classes, the proposed approach can automatically learn which classes are more similar to each other, and construct discriminative features using both labeled and unlabeled data to map similar classes to similar locations in a lower dimensional space. In contrast to linear discriminant analysis (LDA) and its variants, which can only return c-1 features for a problem with c classes, the proposed approach can generate d features, where d is bounded only by the number of the input features. We describe and evaluate the new approach both theoretically and experimentally, and compare its performance with other state of the art methods.

We propose a family of Passive-Aggressive Mahalanobis (PAM) algorithms, which are incremental (online) binary classifiers that consider the distribution of data. PAM is in fact a generalization of the Passive-Aggressive (PA) algorithms to handle data distributions that can be represented by a covariance matrix. The update equations for PAM are derived and theoretical error loss bounds computed. We benchmarked PAM against the original PA-I, PA-II, and Confidence Weighted (CW) learning. Although PAM somewhat resembles CW in its update equations, PA minimizes differences in the weights while CW minimizes differences in weight distributions. Results on 8 classification datasets, which include a real-life micro-blog sentiment classification task, show that PAM consistently outperformed its competitors, most notably CW. This shows that a simple approach like PAM is more practical in real-life classification tasks, compared to more elegant and sophisticated approaches like CW.

NonnegativeMatrix Factorization (NMF) based coclustering methods have attracted increasing attention in recent years because of their mathematical elegance and encouraging empirical results. However, the algorithms to solve NMF problems usually involve intensive matrix multiplications, which make them computationally inefficient. In this paper, instead of constraining the factor matrices of NMF to be nonnegative as existing methods, we propose a novel Fast Nonnegative Matrix Trifactorization (FNMTF) approach to constrain them to be cluster indicator matrices, a special type of nonnegative matrices. As a result, the optimization problem of our approach can be decoupled, which results in much smaller size subproblems requiring much less matrix multiplications, such that our approach works well for large-scale input data. Moreover, the resulted factor matrices can directly assign cluster labels to data points and features due to the nature of indicator matrices. In addition, through exploiting the manifold structures in both data and feature spaces, we further introduce the Locality Preserved FNMTF (LP-FNMTF) approach, by which the clustering performance is improved. The promising results in extensive experimental evaluations validate the effectiveness of the proposed methods.

In this paper, we propose the new Ball Ranking Machines (BRMs) to address the supervised ranking problems. In previous work, supervised ranking methods have been successfully applied in various information retrieval tasks. Among these methodologies, the Ranking Support Vector Machines (Rank SVMs) are well investigated. However, one major fact limiting their applications is that Ranking SVMs need optimize a margin-based objective function over all possible document pairs within all queries on the training set. In consequence, Ranking SVMs need select a large number of support vectors among a huge number of support vector candidates. This paper introduces a new model of of Ranking SVMs and develops an efficient approximation algorithm, which decreases the training time and generates much fewer support vectors. Empirical studies on synthetic data and content-based image/video retrieval data show that our method is comparable to Ranking SVMs in accuracy, but use much fewer ranking support vectors and significantly less training time.

Multi-label learning deals with the problem where each training example is represented by a single instance while associated with a set of class labels. For an unseen example, existing approaches choose to determine the membership of each possible class label to it based on identical feature set, i.e. the very instance representation of the unseen example is employed in the discrimination processes of all labels. However, this commonly-used strategy might be suboptimal as different class labels usually carry specific characteristics of their own, and it could be beneficial to exploit different feature sets for the discrimination of different labels. Based on the above reflection, we propose a new strategy to multi-label learning by leveraging label-specific features, where a simple yet effective algorithm named LIFT is presented. Briefly, LIFT constructs features specific to each label by conducting clustering analysis on its positive and negative instances, and then performs training and testing by querying the clustering results. Extensive experiments across sixteen diversified data sets clearly validate the superiority of LIFT against other well-established multi-label learning algorithms.

Weighted Pairwise Classification (OWPC) loss [Usunier et al., 2009] which has been shown to be state-of-the-art on Image annotation datasets are becoming larger and (small) text retrieval tasks. WARP uses stochastic gradient larger, with tens of millions of images and tens descent and a novel sampling trick to approximate ranks resulting of thousands of possible annotations. We propose in an efficient online optimization strategy which we a strongly performing method that scales to show is superior to standard stochastic gradient descent applied such datasets by simultaneously learning to optimize to the same loss, enabling us to train on datasets that precision at the top of the ranked list of annotations do not even fit in memory.