Due to the difficulties in data collection or data barriers, some crash types may not have sufficient data for analysis, i.e., data scarcity in task (DST). Furthermore, as in crash severity analysis, the number of serious injury or fatality crashes is usually much smaller than that of property damage or injury crashes, there exists class imbalance issue, i.e., data scarcity in class (DSC). To tackle the DST problem, as common crash severity knowledge may hide among cross-task crashes, cross-task crash severity analysis is proposed. Intuitively, by transferring the common crash severity knowledge from source task which contains sufficient crashes to assist the modelling in target task that suffers from DST problem, more effective target task crash severity model can be obtained. Meanwhile, to address the DSC issue, the cost-sensitive learning algorithm for unbiased modelling is adopted by which unequal cost values are assigned to the crashes with different severity levels. In this paper, we develop a cost-sensitive transfer graph convolutional network, where the advanced graph convolutional network is taken as the base learner to guarantee the effectiveness of crash severity analysis. The vehicle-pedestrian and vehicle-vehicle crash severity analysis in city of Seattle are taken as the target and source tasks, respectively. In comparison with eighteen state-of-the-art crash severity models, the proposed model can achieve at least 21.29%, 8.10% and 8.02% improvement in terms of overall summary score, Macro-F1 and Macro-Gmean, respectively. The significant contributing factors analysis also shows that the additional crash contributing factors can be identified, including not only common factors among cross-task data, but also new ground-truth factors of target task crash severity analysis.

This article aims to tackle two issues in hit-and-run crash analysis: The first one is the missing label issue, as the hit-and-run label information has been ignored by many crash datasets. Another one is the class imbalance problem, as the number of hit-and-run crashes is usually much smaller than that of non-hit-and-run crashes. The semi-supervised learning by manually labelling a few hit-and-run labels and the cost-sensitive learning that assigns unequal cost values to class-imbalanced crashes are adopted in this paper. Specifically, we develop a cost-sensitive graph convolutional network with self-paced learning (CSGCNsp) for hit-and-run analysis, which takes the proposed$k$-nearest neighbour anchor graph as input to extract the local clustering structure hidden among crashes for more effective convolutions. The proposed method performs semi-supervised learning with an ‘easy-to-hard’ manner, which can help infer more accurate hit-and-run labels for unsupervised crashes with imbalanced data distribution. The road traffic collision database of Victoria State of Australia is selected for hit-and-run analysis. Compared to fifteen other models, the proposed CSGCNsp can significantly improve the prediction performance in terms of AUC, G-mean, F-measure, and total misclassification cost. The significant contributing factors analysis shows that 53.33% of true contributing factors, which is at least 20.00% higher than that of the state-of-the-art hit-and-run models, can be found by CSGCNsp when only 1% of hit-and-run crashes are manually labelled.

In this article, a cost-sensitive graph convolutional networks (CSGCN) is proposed for crash severity analysis. To extract more complete data information, in addition to the original crash characteristics, CSGCN proposes to take the local neighborhood structure hidden among crashes as the input by defining a$k$-nearest neighbor graph, such that more robust model and significant crash contributing factors can be derived. Furthermore, to address the crash severity class imbalance issue, a cost-sensitive classification layer is developed in CSGCN, aiming to assign unequal cost values to crashes with different crash severity levels. Thereby, an unbiased classification hyperplane can be obtained, especially for estimating the relationship between crash influential factors and the minority but more severe crash severity level. In particular, in order to extract more discriminative crash feature representations, a cost-sensitive graph Laplacian regularisation layer is added to CSGCN. The crash dataset collected from Victorian from 2018 to 2019 is adopted for crash severity analysis. In comparison with 19 crash severity models, the proposed CSGCN can achieve at least 20.91% performance improvement in terms of the designed criterion of overall summary score. Moreover, the pseudo elasticity of the proposed CSGCN shows that the speed zone, accident time, heavy vehicles, road condition, run-off road and collision types are the most significant contributing factors to crash severity, providing implications to traffic engineers and policy makers for crash severity mitigation.

Most existing studies modeled Adaptive Cruise Control (ACC) car-following (CF) behavior using conventional CF models which were originally built for human-driving vehicles (HVs) and calibrated with HV data. In this paper, firstly a learnable CF model is proposed by resorting to Long Short-Term Memory (LSTM) for ACC systems, which utilizes the ACC data for model construction and offers extraordinary adaptability and accuracy. Nevertheless, the applicability of the LSTM CF model is hindered by the scarce ACC data problem, as training the model requires a large amount of data. To address the ACC data scarcity problem, a transfer learning strategy for the LSTM model is further developed, leveraging on the large-scale open-source HV data and the similar driving patterns hidden in HV and ACC-equipped vehicles. In the transfer learning based LSTM model, a unified framework incorporating an alignment layer is developed to transfer the useful features from HV data and meanwhile calibrating the CF model with ACC data. Comparison results show that the proposed model outperforms other CF models which are built with only ACC data or using simple transfer learning methods. Further, microscopic simulations are performed to verify the applicability of the transfer learning based LSTM CF model.

Multi-label hashing is a new research topic in image retrieval. As images are usually associated with multiple semantic labels, there is multi-level semantic similarity such as very similar, normally similar and dissimilar among multi-label images. In order to obtain the multi-level semantic similarity, this letter constructs a hypergraph in label space by creating a hyperedge for each semantic label and including all images annotated with a common label into one hyperedge. In this way, the number of common hyperedges shared by the vertices in hypergraph can be used to encode the high-order semantic relations among multiple images. Considering the useful similarity information hidden in the instance space, a kNN graph in instance space is further constructed. By learning from both the hypergraph and kNN graph with spectral learning strategy, a graph regularized deep discrete hashing is developed which updates graph regularized binary codes and deep neural network based robust features iteratively in a discrete optimization framework. The results in comparison with nine state-of-the-art hashing methods on two multi-label image datasets such as MIRFLICKR-25 K and NUS-WISE demonstrate its effectiveness.

Canonical correlation analysis (CCA) is a cost insensitive method. It assumes the same loss for different classification errors and aims to attain a low error rate by maximizing the cross-view correlation. However, in some real-world applications, different classification errors will lead to unequal misclassification losses. In addition, in practice, only limited cost label information is available in training set due to the expensive costs of labelling. This paper aims to perform label propagation with CCA in a unified cost-sensitive learning framework. By learning jointly, both the label propagation and CCA can feed back to each other. Thus, more discriminative and cost-sensitive projections will be learned for feature fusion. We test the proposed method on the cost-sensitive application of door-locker system based on multi-view face recognition. The results in comparison with eight label propagation methods, eleven CCA related methods and eight cost-sensitive single-view methods demonstrate its effectiveness.

In real-world applications, different kinds of learning and prediction errors are likely to incur different costs for the same system. Moreover, in practice, the cost label information is often available only for a few training samples. In a semi-supervised setting, label propagation is critical to infer the cost information for unlabeled training data. The existing methods typically conduct label propagation independently ahead of supervised cost-sensitive learning. The precomputed label information is kept fixed, which may become suboptimal in the subsequent learning process and hence degrade the overall system performance. In this paper, we develop a unified cost-sensitive framework for semi-supervised face recognition that can jointly optimize the inferred label information and the classifier in an iterative manner. Our experiments on face benchmark datasets demonstrate that in comparison with the state-of-the-art methods for label propagation and cost-sensitive learning, the proposed approach can significantly improve the overall system performance, especially in terms of classification errors associated with high costs.

Similarity search is essential to many important applications and often involves searching at scale on high-dimensional data based on their similarity to a query. In biometric applications, recent vulnerability studies have shown that adversarial machine learning can compromise biometric recognition systems by exploiting the biometric similarity information. Existing methods for biometric privacy protection are in general based on pairwise matching of secured biometric templates and have inherent limitations in search efficiency and scalability. In this paper, we propose an inference-based framework for privacy-preserving similarity search in Hamming space. Our approach builds on an obfuscated distance measure that can conceal Hamming distance in a dynamic interval. Such a mechanism enables us to systematically design statistically reliable methods for retrieving most likely candidates without knowing the exact distance values. We further propose to apply Montgomery multiplication for generating search indexes that can withstand adversarial similarity analysis, and show that information leakage in randomized Montgomery domains can be made negligibly small. Our experiments on public biometric datasets demonstrate that the inference-based approach can achieve a search accuracy close to the best performance possible with secure computation methods, but the associated cost is reduced by orders of magnitude compared to cryptographic primitives.

In recent years, face recognition is being recognized as a cost-sensitive learning problem. Many cost-sensitive classifiers have been proposed. However, no sufficient attention is paid to the research on cost-sensitive dimensionality reduction, especially on the cost-sensitive semisupervised dimensionality reduction. To the best of our knowledge, cost sensitive semisupervised discriminant analysis (CS3DA) may be the first work. CS3DA first uses the sparse representation to infer a soft label for unlabeled sample and then learns the projection direction by incorporating misclassification costs into both labeled and unlabeled data. Although CS3DA reduces the loss of misclassification, it has two major drawbacks: 1) the sparsity is not a feature of face recognition, and therefore sparse approximations may not deliver the robustness or performance desired and 2) CS3DA is not proven to satisfy the minimal misclassification loss criterion. In this paper, we embed pairwise costs in semisupervised discriminant analysis (PCSDA) for face recognition. PCSDA first uses a simple l2approach to predict the label of unlabeled data, and then learns the projection direction by embedding pairwise costs in both labeled and unlabeled data. Compared with CS3DA, PCSDA has three major advantages: 1) l2approach is more accurate and robust than sparse representation for face recognition; 2) we prove that CS3DA approximates the pairwise Bayesian risk only when the classes are balanced and without outliers in face data sets; and 3) PCSDA approximates the pairwise Bayesian risk considering the class imbalance problem and outliers in face recognition. Hence, the projection direction obtained by using PCSDA can be more discriminative, immunes to outliers and class imbalance problem. The experimental results on AR, PIE, ORL, and extended Yale B data sets demonstrate the effectiveness of PCSDA.