Joint label-specific features and label correlation for multi-label learning with missing label

Feature selection for multilabel classification with missing labels via multi-scale fusion fuzzy uncertainty measures

Facial action unit (AU) recognition has attracted great attention because of the applications in a wide range of fields. Missing labels and class imbalance (CIB) are both challenges for facial action unit recognition. Missing labels means that there are only apart label assignments for training samples. CIB is observed from two perspectives: firstly, the number of positive AUs is much smaller than that of negative AUs for each expressional image; secondly, the rate of positive samples of different AUs are significantly different. Both missing labels and CIB lead to performance degradation in AU recognition. In this work, we propose to handle these two challenges in AU recognition simultaneously. Specifically, we formulate AU recognition with missing labels as a multi label learning with missing labels (MLML) problem, which handles the missing label challenge naturally. However, different from most existing MLML approaches which usually employ same features from whole image for all classes, we select the most related features for each AU. To handle the CIB challenge, we further introduce class cardinality bounds which constrain the number of positive AUs for each data instance, as well as the number of positive labels for each AU in the overall dataset. The class cardinality bounds serve as linear constraints for the objective function, which turns the optimization NP-hard. Thus we present convex approximation based on the Lovasz extension, which leads to a linear program that can be efficiently solved by the alternative direction method of multipliers (ADMM). Experimental results on both posed and spontaneous facial expression datasets demonstrate the superiority of the proposed method compared to state-of-the-art.

To deal with the problem where each instance is associated with multiple labels, a lot of multi-label learning algorithms have been developed in recent years. Some approaches have been proposed to select label-specific features to utilize discriminate features for multi-label classification. Although label correlation has been considered in learning label-specific features, the critical correlation among instances was less taken into account. In this paper, we proposed a new approach called multi-label learning with label-specific features using correlation information (LSF-CI) to learn label-specific features for each label with the consideration of both correlation information in label space and correlation information in feature space. In the LSF-CI, the instance correlation in feature space is computed by a probabilistic neighborhood graph model, and label correlation in label space is computed by cosine similarity. For multi-label data, the LSF-CI has the capability to select Label-specific features for each label as well as classify an unseen instance into a set of relevant labels. To validate the effectiveness of LSF-CI, we conducted comprehensive experiments on eight multi-label datasets. The experimental results demonstrate that the LSF-CI is capable of selecting compact label-specific features, and achieving a competitive performance in comparison with the performances of the existing multi-label learning approaches.

In multi-label learning with missing labels, previous works usually only consider all features, or only consider label-specific features. It is obviously inappropriate to select a subset of features only considering which have a great discriminability for all labels or for a label. Besides, they are built on an assumption that if two labels are correlated, their regression coefficient vectors should be similar. However, it is hard to hold this assumption in real-world applications. Therefore, we propose a novel multi-label classification approach for missing labels via considering common and label-specific features with correlation information. More specifically, we firstly utilize <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$l_{1}$</tex> - norm regularizer over label correlation matrix, and a new sup-plementary label matrix is augmented from the incomplete label matrix by learning label correlations. Then, for a better label recovery, we introduce <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$l_{2,1}$</tex> -norm and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$l_{1}$</tex> -norm regularizers to learn common and label-specific features simultaneously. Thirdly, to make more rational use of label correlation to solve missing labels, we use a regularizer to constrain label correlations on outputs of labels instead of regression coefficient matrix. Finally, an objective function is designed in terms of the above processing. Extensive experiments conducted on eight data sets demonstrate the effectiveness of the proposed approach.

Existing multi-label learning algorithms mostly assume that all class labels in an obtained data set are completely accurate. This assumption is rarely valid. As the scale of the label space increases, owing to limited domain knowledge and differences among methods of data acquisition, some examples will be unlabeled or mislabeled. The introduction of such noise results in the poor performance of trained multi-label classifiers. This paper proposes a new method for multi-label learning with label noise, called low-rank-constraint-based multi-label learning with label noise (ML-LRC), by incorporating low-rank matrix recovery and features selection into the learning process. First, we assume that two strongly correlated labels share more samples than two weakly correlated labels. Based on this assumption, the ML-LRC applies low-rank constraints to the label matrix to mine the local correlations of class labels and remove noise from the observed label matrix. Second, we believe that each class label only correlates to some features. So we control the sparsity of the coefficient matrix to filter out label-specific features, which can reduce the complexity of the model and greatly improve the performance of the classifier. The accelerated proximal gradient method and alternating direction method of multipliers are adopted to solve the problem. Experiments conducted on four benchmark multi-label datasets demonstrate the competitive performance of our proposed method compared with several state-of-the-art methods.

Multi-label learning has attracted significant interests in computer vision recently, finding applications in many vision tasks such as multiple object recognition and automatic image annotation. Associating multiple labels to a complex image is very difficult, not only due to the intricacy of describing the image, but also because of the incompleteness nature of the observed labels. Existing works on the problem either ignore the label-label and instance-instance correlations or just assume these correlations are linear and unstructured. Considering that semantic correlations between images are actually structured, in this paper we propose to incorporate structured semantic correlations to solve the missing label problem of multi-label learning. Specifically, we project images to the semantic space with an effective semantic descriptor. A semantic graph is then constructed on these images to capture the structured correlations between them. We utilize the semantic graph Laplacian as a smooth term in the multi-label learning formulation to incorporate the structured semantic correlations. Experimental results demonstrate the effectiveness of the proposed semantic descriptor and the usefulness of incorporating the structured semantic correlations. We achieve better results than state-of-the-art multi-label learning methods on four benchmark datasets.

Local positive and negative correlation-based k-labelsets for multi-label classification

Multi-label classification with local pairwise and high-order label correlations using graph partitioning

In multi-label learning, an efficient approach with label-specific features named LIFT has been presented, since different labels may have some distinct characteristics. However, the construction of label-specific features by simply assigning equal weight to each instance ignores the relevance among samples, which might increase the dimensionalities and result in a large amount of redundant information. In order to reduce the redundancy, a novel yet effective multi-label learning approach with weighted label-specific feature selection by using information theory (WFSI-LIFT) is proposed. In WFSI-LIFT, we employ the information theory to implement label-specific feature selection and assign different weights to the different class instance according to imbalance rate(IR). And then, comprehensive experiments across 8 real-world multi-label data sets indicate that, WFSI-LIFT can not only reduce the dimensionalities of label-specific features and enhance the performance compared with LIFT, but also validate the superiority of our approach against other well-established multi-label learning algorithms.

Partial multi-label learning via specific label disambiguation

The choice of the loss function is critical in extreme multi-label learning where the objective is to annotate each data point with the most relevant subset of labels from an extremely large label set. Unfortunately, existing loss functions, such as the Hamming loss, are unsuitable for learning, model selection, hyperparameter tuning and performance evaluation. This paper addresses the issue by developing propensity scored losses which: (a) prioritize predicting the few relevant labels over the large number of irrelevant ones; (b) do not erroneously treat missing labels as irrelevant but instead provide unbiased estimates of the true loss function even when ground truth labels go missing under arbitrary probabilistic label noise models; and (c) promote the accurate prediction of infrequently occurring, hard to predict, but rewarding tail labels. Another contribution is the development of algorithms which efficiently scale to extremely large datasets with up to 9 million labels, 70 million points and 2 million dimensions and which give significant improvements over the state-of-the-art. This paper's results also apply to tagging, recommendation and ranking which are the motivating applications for extreme multi-label learning. They generalize previous attempts at deriving unbiased losses under the restrictive assumption that labels go missing uniformly at random from the ground truth. Furthermore, they provide a sound theoretical justification for popular label weighting heuristics used to recommend rare items. Finally, they demonstrate that the proposed contributions align with real world applications by achieving superior clickthrough rates on sponsored search advertising in Bing.

Multi-label image classification aims to predict the labels associated with a given image. While most existing methods utilize unified image representations, extracting label-specific features through input space learning would improve the discriminative power of the learned features. On the other hand, most feature learning studies often ignore the learning in the output label space, although taking advantage of label correlations can boost the classification performance. In this paper, we propose a deep learning framework that incorporates flexible modules which can learn from both input and output spaces for multi-label image classification. For the input space learning, we devise a label-specific feature pooling method to refine convolutional features for obtaining features specific to each label. For the output space learning, we design a Two-Stream Graph Convolutional Network (TSGCN) to learn multi-label classifiers by mapping spatial object relationships and semantic label correlations. More specifically, we build object spatial graphs to characterize the spatial relationships among objects in an image, which supplements the label semantic graphs modelling the semantic label correlations. Experimental results on two popular benchmark datasets (i.e., Pascal VOC and MS-COCO) show that our proposed method achieves superior performance over the state-of-the-arts.

In a conventional supervised learning paradigm, each data instance is associated with one single class label. Multilabel learning differs in the way that data instances may belong to multiple concepts simultaneously, which naturally appear in a variety of high impact domains, ranging from bioinformatics and information retrieval to multimedia analysis. It targets leveraging the multiple label information of data instances to build a predictive learning model that can classify unlabeled instances into one or multiple predefined target classes. In multilabel learning, even though each instance is associated with a rich set of class labels, the label information could be noisy and incomplete as the labeling process is both time consuming and labor expensive, leading to potential missing annotations or even erroneous annotations. The existence of noisy and missing labels could negatively affect the performance of underlying learning algorithms. More often than not, multilabeled data often has noisy, irrelevant, and redundant features of high dimensionality. The existence of these uninformative features may also deteriorate the predictive power of the learning model due to the curse of dimensionality. Feature selection, as an effective dimensionality reduction technique, has shown to be powerful in preparing high-dimensional data for numerous data mining and machine-learning tasks. However, a vast majority of existing multilabel feature selection algorithms either boil down to solving multiple single-labeled feature selection problems or directly make use of the imperfect labels to guide the selection of representative features. As a result, they may not be able to obtain discriminative features shared across multiple labels. In this article, to bridge the gap between a rich source of multilabel information and its blemish in practical usage, we propose a novel noise-resilient multilabel informed feature selection framework (MIFS) by exploiting the correlations among different labels. In particular, to reduce the negative effects of imperfect label information in obtaining label correlations, we decompose the multilabel information of data instances into a low-dimensional space and then employ the reduced label representation to guide the feature selection phase via a joint sparse regression framework. Empirical studies on both synthetic and real-world datasets demonstrate the effectiveness and efficiency of the proposed MIFS framework.

On social media, the user generated contents, e.g., Articles and images, can be assigned with multiple labels. In this paper, we focus on the problem of performing multi-label classification on social media data, where the user generated contents are associated with multiple labels. Multi-label learning studies the problem where each object is represented by a single instance and associated with a set of labels. Current multi-label learning algorithms mainly exploit label correlations globally, by assuming that the label correlations are shared by all the examples. In real applications, however, different examples may share different label correlations. In this paper, we propose a Local Pair wise Label Correlation (LPLC) method for social media content categorization. We try to exploit the strongest local pair wise label correlations between the ground truth labels for each training example by computing the maximum conditional probabilities. If two labels have strong correlation, there will be a larger conditional probability of one label given by another. In the training stage, we find the most correlated labels for ground truth labels of each training example. In the test stage, we make prediction through maximizing the posterior probability, which is estimated with the distribution of each label in the k nearest neighbors and their most correlated local pair wise label correlations. We compare our method with six well-established multi-label learning algorithms over nine data sets from different social media data domains and scales. Comparison results with the state-of-the-arts approaches manifest competitive performances of our method.

In multi-label learning, research on label correlation provides an effective solution to compress the hypothesis space of classifiers. However, the current works focus on the label correlation adapted to overall data, while ignoring the locally targeted information presented by some instances. The lack exploration on the distribution of local label correlation in multi-label instance space undoubtedly limits the in-depth application of label correlation in multi-label learning. Based on fuzzy rough set theory, the instance-level label correlation distribution is first proposed in this paper and applied to design a novel multi-label learner. For each multi-label instance, the local importance of features to label is quantitatively analyzed, by considering the decisive influence of input information on decision-making. According to coincidence degree between local feature weight distribution for different labels, the instance-level label correlation is constructed. In order to reflect the internal relationship between label variables objectively, the instance-level label correlation distribution is integrated into the empirical label relevance. On the basis, the label relevance matrix is used to define the constraints of optimization function in a new form. The relative position of sub-separating hyperplanes in input space is quantitatively characterized to reduce the complexity of multi-label classifier and improve learning performance. The experiment results on eighteen multi-label datasets illustrate the effectiveness of our algorithm. The impact of core parameters on performance is also dissected.