Objective.This study aimed to employ a two-stage deep learning method to accurately detect small aneurysms (4-10 mm in size) in computed tomography angiography images.Approach.This study included 956 patients from 6 hospitals and a public dataset obtained with 6 CT scanners from different manufacturers. The proposed method consists of two components: a lightweight and fast head region selection (HRS) algorithm and an adaptive 3D nnU-Net network, which is used as the main architecture for segmenting aneurysms. Segments generated by the deep neural network were compared with expert-generated manual segmentation results and assessed using Dice scores.MainResults.The area under the curve (AUC) exceeded 79% across all datasets. In particular, the precision and AUC reached 85.2% and 87.6%, respectively, on certain datasets. The experimental results demonstrated the promising performance of this approach, which reduced the inference time by more than 50% compared to direct inference without HRS.Significance.Compared with a model without HRS, the deep learning approach we developed can accurately segment aneurysms by automatically localizing brain regions and can accelerate aneurysm inference by more than 50%.

Recently, query-based instance segmentation methods have achieved comparable performance to previous state-of-the-art methods. However, the query lacks the learning of the consistency between classification and segmentation tasks, which may lead to misalignment between classification score and mask quality (i.e., mask IoU) and can not result in a reliable ranking for predictions. In this work, we propose a novel instance segmentation method, termed AlignMask, which effectively learns task-aligned mask queries for instance end-toend. Specifically, we propose Aligned Query Learning (AQL) to learn task-aligned features for pixel embedding and transformer decoder, which helps segmentation quality estimation of the mask query. We also use Aligned Label Assignment to explicitly align the optimization goals for classification score and mask quality of the query. Extensive experiments on MSCOCO show that our proposed AlignMask achieves competitive performance with state-of-the-art models.

Person-Job Fit (PJF) is the core of online recruitment. Several recent methods took PJF as a preference-learning problem, and tried to learn two-sided preferences from their historical behaviors. However, they ignored users' interactive feedbacks (accepted or rejected) received from the other side which may change users' preferences. In addition, they neglected the status of local market which may affect the final matching results of person-job pairs. To solve these issues, we propose a market-aware dynamic PJF method with hierarchical reinforcement learning (HIRE). We design a two-level hierarchy of reinforcement learning policies. Two low-level policies aim to learn the dynamic preferences of both persons and jobs with consideration of interactive feedbacks, and a high-level policy aims to learn the optimal dynamic matching strategy with consideration of local market state. Extensive experiments on two real-world datasets show the effectiveness of HIRE compared with the state-of-the-art.

Data sparsity is a common problem in collaborative ranking for personalized recommendation with implicit feedback. Several previous work tried to ‘borrow’ feedback information from users’ neighborhood as their prior preferences to alleviate this problem. However, they emphasize the overlapping interests of users and their neighborhood while de-emphasize the importance of users’ own specific taste, which leads to under-personalization. In addition, they ignore the collaborative influence among items which is also important for preference learning.To solve these problems, we propose an effective collaborative ranking method G eneralized C ollaborative P ersonalized R anking (GCPR), which utilizes the collaborative influence among users and items in a unified framework. It strengthens the specific taste of users by using inner-basket influence of items to enhance the personalization. In addition, it utilizes cross-basket influence of items to dig more collaborative items to further alleviate the sparsity problem. Then, we utilize generalized AUC to learn a confidence-based listwise preference, and propose a post-training based on self-paced learning to solve the top-biased problem of the generalized AUC. Experimental results on four public real-world datasets show that GCPR achieves better performance than traditional collaborative filtering (CF) methods and state-of-the-art collaborative ranking methods.

Recommendation based on heterogeneous information network(HIN) is attracting more and more attention due to its ability to emulate collaborative filtering, content-based filtering, context-aware recommendation and combinations of any of these recommendation semantics. Random walk based methods are usually used to mine the paths, weigh the paths, and compute the closeness or relevance between two nodes in a HIN. A key for the success of these methods is how to properly set the weights of links in a HIN. In existing methods, the weights of links are mostly set heuristically. In this paper, we propose a Bayesian Personalized Ranking(BPR) based machine learning method, called HeteLearn, to learn the weights of links in a HIN. In order to model user preferences for personalized recommendation, we also propose a generalized random walk with restart model on HINs. We evaluate the proposed method in a personalized recommendation task and a tag recommendation task. Experimental results show that our method performs significantly better than both the traditional collaborative filtering and the state-of-the-art HIN-based recommendation methods.

Error decomposition analysis is a key problem for ensemble learning. Two commonly used error decomposition schemes, the classic Ambiguity Decomposition and Bias-Variance-Covariance decomposition, are only suitable for regression tasks with square loss. We generalized the classic Ambiguity Decomposition from regression problems with square loss to classification problems with any loss functions that are twice differentiable, including the logistic loss in Logistic Regression, the exponential loss in Boosting methods, and the 0-1 loss in many other classification tasks. We further proved several important properties of the Ambiguity term, armed with which the Ambiguity terms of logistic loss, exponential loss and 0-1 loss can be explicitly computed and optimized We further discussed the relationship between margin theory, "good" and "bad" diversity theory and our theoretical results, and provided some new insights for ensemble learning. We demonstrated the applications of our theoretical results in active learning and unsupervised ensemble pruning, and the experimental results confirmed the effectiveness of our methods.

Discretization and feature selection are two basic preprocessing stages of data mining. However, it often results in information loss due to these two separate stages. This paper proposes a novel supervised multivariate discretizer integrated with feature selection, called D-FS. It takes into consideration of the interactions of both different cut-points and features, and achieves feature selection by discretization. D-FS can avoid the information loss caused by the independence of discretization and feature selection. Compared with several state-of-the-art discretizers, D-FS retains a smaller subset of both cut-points and features, while achieves competitive classification performance combined with different classifiers.

The rise of online P2P lending, as a novel economic lending model, brings new opportunities and challenges for the research of credit risk evaluation. This paper aims to mine information from different data sources to improve the performance of credit risk evaluation models. Be-sides the personal financial and demographic data used in traditional models, the authors collect in-formation from (1) text description, (2) social network and (3) macro-economic data. They de-sign methods to extract features from unstructured data. To avoid the curse of dimensionality caused by too many features and identify the key factors in credit risk, the authors remove the irrelevant and redundant features by feature selection. Using the data provided by Prosper.com, one of the biggest P2P lending platforms in the world, they show that: (1) it can achieve better performance, measured by both AUC (area under the receiver operating characteristic curve) and classification accuracy, by fusion of information from different data sources; (2) it requires only ten features from different data sources to get better performance.

Feature selection is an important step for data mining and machine learning to deal with the curse of dimensionality. In this paper, we propose a novel feature selection framework based on supervised data clustering. Instead of assuming there only exists low-order dependencies between features and the target variable, the proposed method directly estimates the high-dimensional mutual information between a candidate feature subset and the target variable through supervised data clustering. In addition, it can automatically determine the number of features to be selected instead of manually setting it in a prior. Experimental results show that the proposed method performs similar or better compared with state-of-the-art feature selection methods.

In ensemble learning, ensemble pruning is a procedure that aims at removing the unnecessary base classifiers and retaining the best subset of the base classifiers. We presented a two-step ensemble pruning framework, in which the optimal size of the pruned ensemble is first decided, and then with the optimal size as input, the optimal ensemble is selected. For the first step to find the optimal ensemble size, we presented an algorithm that can be proved to be able to find the Bayesian optimal ensemble size. For the second step, we developed two greedy forward pruning methods, i.e., the Bayesian Pruning (BP) method and the Bayesian Independent Pruning (BIP) method. In the BP method, we assumed that the probability of a candidate ensemble to be the optimal ensemble follows the Generalized Beta distribution. And in the BIP method, we further assumed that whether a base classifier belongs to the optimal ensemble is independent to the other base classifiers. Experimental results on twenty benchmark data sets showed that the BP and BIP methods achieved competitive performance in contrast to other state-of-the-art algorithms.

In this paper, we propose a coordinate transform model and document rectification process for book dewarping. This model can handle both perspective distortion and book surface warping problems. The main idea is to find the boundary of a region, which is a rectangle on a flat surface before warping and using the correspondence between the warped boundary and the original rectangle sides to map each pixel in the warped region into the rectangle. On a book page, two text lines and the left and right boundaries of a text region form a rectangle before warping. Therefore, we detect two text flow curves and text region boundaries on the left and right sides on a warped image and use the two curves and two line segments to do the transformation. The experimental results illustrate the effectiveness of our proposed book dewarping approach.

Fiber reinforced composites are applied broadly in aeronautic and astronautic fields as a structural material. But the investigation in dynamic fracture behavior of fiber reinforced composite stands in the breach for scientists due to a large number of aircraft disasters. In this paper, the mixed mode fracture problems in fiber reinforced composites under impact are studied. First, based on the theory of the reflective dynamic caustic method for mixed mode fracture, corresponding experiments are carried out to study the dynamic fracture behaviors of unidirectional fiber reinforced composites under two kinds load conditions. By recording and analyzing the shadow spot patterns during the crack propagation process carefully, the dynamic fracture toughness and crack growth velocity of fiber reinforced composites are obtained. Via the observation of the crack growth routes and fracture sections, we further reveal the fracture mechanism of unidirectional fiber reinforced composites. It concludes that opening mode still is the easier fracture type for the pre-crack initiation in fiber reinforced composites, while the interface between fibers and matrix becomes the fatal vulnerability during the crack propagation.

In this paper, we propose a book dewarping model based on the assumption that the book surface is warped as a cylinder. This model extends the model proposed by Cao and makes Cao’s model a special case of our model. This extension removes the constraint of Cao’s model that the camera lens must be strictly parallel to the book surface, which is hard to make in practice, therefore enables a user to take a picture from different point of views conveniently. The main idea of the model is to build up the correspondence between a rectangle region on a flat surface and its curved region on the distorted book image and the dewarping task is to flatten the curved region to its original rectangle shape. The experimental results demonstrate the effectiveness of our proposed book dewarping approach.

In this paper, we propose a book dewarping model which flattens a curved book page to its original flat rectangle shape. This model generalizes the model proposed by Cao to handle warped book images taken from different point of views instead of only from top of the book surface. In order to do so, we realize the impact of the angle between the camera lens and the book surface and take it into account in the transform model. Based on the new model, the rectification process includes two steps: 1) on the warped book image, find a region which is original a rectangle on a flat page; 2) map each pixel in warped region to a pixel in its original rectangle region. The experimental results demonstrate the effectiveness of our proposed book dewarping approach.

The microstructures and intrinsic defects engendered during the compression process of manufacture of a PBX substitute material D-90031, as well its dynamic behavior and failure mechanisms were studied. By using digital image treatment technique, in-situ nano-indenter system, Split Hopkin-son Pressure Bar (SHPB) and weight dropping tests, we obtained the features of microstructures, modulus and hardness of each constituent, dynamic properties, as well as the size distribution of pieces and failure modes. It was found that the grain shape changed with the applied compressive load, such as the grain was elongated along the direction vertical to the load application direction. The size distribution of pieces after impact meets a certain statistical rule. The results obtained are meaningful to deeply understand the relationship between the properties and microstructures of solid explosive.

In this paper, we propose a coordinate transform model to rectify both the perspective and surface curving distortions of a book image, which is captured by a 2D digital camera. The rectification process includes three steps: 1) detect the boundary of a curved page on a book image; 2) knowing the four boundary curves of a book page should be mapped to the four sides of a rectangle, we build up the correspondence between a pixel inside the book page and a pixel inside the rectangle; 3) fill each pixel in the rectangle with its correspondence pixel on the book image, we get a rectangle which is the dewarped page we want. The experimental results show the effectiveness of the proposed rectification approach.