MLSLR: آموزش برچسب چندگانه از طریق پراکنده رگرسیون لجستیک
|کد مقاله||سال انتشار||مقاله انگلیسی||ترجمه فارسی||تعداد کلمات|
|25004||2014||11 صفحه PDF||سفارش دهید||7073 کلمه|
Publisher : Elsevier - Science Direct (الزویر - ساینس دایرکت)
Journal : Information Sciences, Volume 281, 10 October 2014, Pages 310–320
Multilabel learning, an emerging topic in machine learning, has received increasing attention in recent years. However, how to effectively tackle high-dimensional multilabel data, which are ubiquitous in real-world applications, is still an open issue in multilabel learning. Although many efforts have been made in variable selection for traditional data, little work concerns variable selection for multilabel data yet. In this paper, we propose a novel framework for multilabel learning, which can achieve the purposes of variable selection and classification learning simultaneously. Specifically, our method exploits logistic regression to train models on multilabel data for classification. Besides, an elastic net penalty is performed on the logistic regression model to handle ill-conditioned and over-fitting problems of high-dimensional data. To further improve the efficiency, we solve the convex optimization problem of logistic regression with the elastic net penalty by a quadratic approximation technique. The experimental results on seven multilabel data sets show that our method has achieved encouraging performance and is competitive with six popular multilabel learning algorithms in most cases.
Supervised learning is a major task in data mining and machine learning and has been extensively studied during past four decades. Traditional supervised learning methods place emphases on the data (or samples), which are associated with class labels exclusively. However, in many real-world applications objects can often be tagged with two or more class labels simultaneously. For example, the movie ‘avatar’ may be tagged with action, science fiction and love types; The journal of Information Sciences is associated with journal, Elsevier, computer science and so on. This kind of data is called multilabel data . Multilabel data are ubiquitous in many domains, such as text categorization, image annotation and bioinformatics  and , and can be regarded as a special representation of multi-view data . Multilabel learning has attracted significant attention from many interdisciplinary fields since it was introduced , because multilabel learning has a large number of potential applications in reality. By now, many multilabel learning algorithms have been developed and successfully applied in text categorization , images and video annotation , content annotation , music processing , bioinformatics , and so on. Generally, the multilabel learning algorithms can be categorized into two major groups, i.e., problem transformation and algorithm adaption . Comparing to traditional learning methods, the characteristic of multilabel learning is that the outputs of multilabel learning methods are not mutually exclusive and they may even be correlated with each other in some cases. A challenging issue for multilabel learning is the problem of ‘curse of dimensionality’ raised from the high-dimensional multilabel data. With advancements of information techniques, data in many fields are turning to large in both size and dimensionality. As the dimensionality of data is getting larger, the so-called problem of multi-collinearity may occur with a high probability, resulting in multilabel learning more challenging and complicated. An effective solution of dealing with this troublesome issue is to perform dimension reduction on the multilabel data before constructing learning models. For example, Zhang and Zhou  projected the original data into a lower-dimensional space with the Hilbert–Schmidt Independence Criterion, while Ji et al.  extracted a common subspace shared among multiple labels by virtue of ridge regression. It is noticeable that this kind of work merely focuses on extracting common subspaces with fewer dimensions, but pays little attention on the interpretability of results. In fact, the derived subspaces are weighted combinations of the original ones by linear transformation, making the interpretation of results impossible . In this paper, we investigate the problems of variable selection and classification learning for multilabel data, and then propose a joint multilabel learning framework. It can fulfill the purposes of multilabel classification and variable selection simultaneously. To the best of our knowledge, variable selection in the context of multilabel learning has not been fully exploited yet and is still an open issue, albeit it has been well investigated in the traditional learning methods . For multilabel learning, there is a variety of reasons to perform variable selection, such as lessening computational cost, avoiding over-fitting, and improving prediction performance and interpretability  and . To achieve variable selection, we impose an elastic net penalty on logistic regression for multilabel learning, where the ℓ1ℓ1-norm penalty of the elastic net aims at removing irrelevant variables, while the ℓ2ℓ2-norm penalty ensures that highly correlated variables have similar regression coefficients . The main contributions of this paper are as follows: • We propose a general framework of learning and variable selection for multilabel data, which can conduct classification learning and variable selection simultaneously. The purpose of classification learning is achieved via logistic regression, which is particularly capable of solving the problem of binary classification. • We perform the elastic net penalty to the logistic regression model, where the ℓ1ℓ1-norm penalty provides a solution for variable selection, yielding a sparse model, while the ℓ2ℓ2-norm penalty offers a grouping effect of the correlated variables, and a unique solution when the number of variables is larger than samples. • To further improve learning efficiency, we explore the convex optimization problem of the sparse logistic regression by virtue of a quadratic approximation technique. Paper organization. The rest of this paper is organized as follows. Section 2 briefly reviews the state-of-the-art of dimension reduction techniques for multilabel data. The problem of multilabel learning is formulated in Section 3. Section 4 gives the model of our sparse logistic regression, and provides an analytical solution to this optimization problem. We report the experimental results on seven data sets in Section 5, and then conclude the paper in Section 6. Notations. Throughout this paper, uppercase bold Roman letters denote matrices and lowercase ones denote column vectors. For a specific example, xx is a column vector while XX represents the matrix [x1,…,xd][x1,…,xd]. [·]T[·]T indicates the transpose of a matrix or vector. ‖·‖k‖·‖k denotes the Frobenius norm for matrices and k norm for vectors.
نتیجه گیری انگلیسی
In this paper, we proposed a framework of classification learning and variable selection for multilabel data. According to the characteristics of multilabel data, a logistic regression model was exploited in this framework for classification learning. To cope with the ill-conditioned and over-fitting problems from the high-dimensional data, the elastic net penalty was imposed on the logistic regression model. Thus, the purposes of variable selection and group effects can be achieved simultaneously. Experimental results on seven multilabel data sets have demonstrated that the proposed method can potentially improve performance and outperform competing methods in most cases. Currently, we are only concerning about the linear dependency of variables in our method. In the near future, we will investigate nonlinear dependency of variables and place our focus on modeling nonlinear dependency to extend the proposed method with suitable kernel functions for more complex applications.