Vol. 44 · No. 1

According to Hughes phenomenon, the major challenges encountered in computations with learning models come from the scale of complexity, e.g. the so-called curse of dimensionality. Approaches for accelerated learning computations range from model- to implementation-level. The first type is rarely used in its basic form. Perhaps, this is due to the theoretical understanding of mathematical insights. We describe a model-level decomposition approach that combines both the decomposition of the objective function and of data. We perform a feasibility analysis of the resulting algorithm, both in terms of accuracy and scalability.

Analysis class diagram (ACD) metrics like number of classes, number of methods, and number of attributes can be used for early software size estimation by project managers during initial project planning. However, not all of these ACD metrics have the same influence on software size. This study aims to empirically determine the relative influence of these ACD metrics on software size using historical data from academia and industry. Using the objective class points (OCP) metric as a base, two new metrics – enhanced OCP (EOCP) and weighted EOCP (WEOCP) – are proposed. Separate linear regression-based early software size estimation models are also constructed and validated using the original OCP metric and its newly proposed variants. A comparison of these models reveals that models based on our freshly proposed metrics perform better in terms of early size estimation accuracy.

Community structure is a common feature of social networks and many community discovery algorithms have emerged through the study of this feature. The gravitational field model is an effective method to realize community division. However, the current gravitational field model lacks a comprehensive consideration of field properties such as the internal stability of the gravitational field. Therefore, in this paper, we define and quantify the attributes of the gravitational field by taking advantage of the field's strength in describing the joint action of groups. Then, we propose a social network gravitational field community detection model (GF-CDM). GF-CDM selects the field kernel node based on a random walk and then presents an adaptive expansion function of fusion field stability to divide the observable network into overlapping and non-overlapping clusters. The model was evaluated on four real network datasets and five artificial network datasets of different sizes. Experimental results show that our proposed model outperforms the other four benchmark algorithms in modularity, ARI index, and field average stability, which can improve the quality of cluster division.

Semantic segmentation is one of the key technologies in the development of autonomous vehicles. Practical applications are increasingly pursuing a balance between effectiveness and efficiency. Many lightweight segmentation models nowadays have some problems, often making it difficult to predict small objects and edges between different objects. In this work, we propose a model of encoder-decoder structure, DeliteSeg. Firstly, we added deformable convolutional layers to the encoder, leveraging the advantages of deformable convolution to enable the model to better predict object edges. Then we proposed a new deep context aggregation module DLPPM, which improves the context information aggregation ability by fusing low-resolution feature maps of different scales multiple times, enabling the model to better predict small objects. Finally, we designed a new lightweight attention decoder (LMD) that utilizes a spatial channel attention mechanism to refine feature maps at different levels, effectively recovering information. After extensive experiments, our network achieved 73.6

Few-shot semantic segmentation is a challenging task aimed at segmenting new objects in the query image with only a few annotated support images. Most advanced methods for this task mainly focus on either global or local prototype learning through global average pooling (GAP) or clustering. However, due to the limitation of average and cluster operation, these methods still fail to exploit the object information from support images entirely. To address these limitations, we propose a generalization of prototype learning in the frequency domain through multi-frequency pooling (MFP) to mine both local and global object information. Based on the MFP, we further build a Frequency Prototype Network (FPNet) consisting of three novel designs. Firstly, the Frequency Prototype Generation Module (FPGM) extracts frequency prototypes by MFP in the DCT domain to provide complete object guidance information. Then, the Prior Attention Mask Module (PAMM) produces a prior attention mask to identify a query target more precisely and retain high generalization. Finally, the Frequency Prototype Selection Module (FPSM) selects the most effective support prototypes to reduce redundancy. Extensive experiments on PASCAL-5i and COCO-20i demonstrate that our model achieves state-of-the-art performances in both 1-shot and 5-shot settings.

Breast cancer (BC) is one of the leading causes of death in women worldwide. Early diagnosis of this disease can save many women's lives. The Breast Imaging Reporting and Data System (BIRADS) is a standard method developed by the American College of Radiology (ACR). However, physicians have had a lot of contradictions in determining the value of BIRADS, and all aspects of patients have not been considered in diagnosing this disease using the methods that have been used so far. In this article, a novel decision support system (DSS) has been presented. In the proposed DSS, firstly, c-mean clustering was used to determine the molecular subtype for patients who did not have this value by combining the mammography reports processing along with hospital information systems (HIS) obtained from their electronic files. Then several classifiers such as convolutional neural networks (CNN), decision tree (DT), multi-level fuzzy min-max neural network (MLF), multi-class support vector machine (SVM), and XGboost were trained to determine the BIRADS. Finally, the values obtained by these classifiers were combined using ensemble learning with the majority voting algorithm to obtain the appropriate value of BIRADS. This helps physicians in the early diagnosis of BC. Finally, the results were evaluated in terms of accuracy, specificity, sensitivity, positive predicted value (PPV), negative predicted value (NPV), f1-measure, and balanced accuracy by the confusion matrix. The obtained values were 87.77

In the complex environment of industrialized brown mushroom cultivation, a dense brown mushroom detection algorithm based on improved YOLOv7 is proposed to address the issues of low real-time detection accuracy and speed, and the high false detection rate of picking robots in densely grown brown mushroom clusters. To prevent network degradation, improve the detection accuracy and speed of the network, and reduce the network's computational cost, the ELAN PS module is introduced to replace the original ELAN module. The AFPN network is used to replace the original network's Neck part for multi-scale fusion, allocating different spatial weights to feature maps to enhance the model's ability to separate dense targets. The MDIoU loss function is introduced as the algorithm's bounding box loss function to optimize the convergence speed of network training and improve the detection accuracy of dense occluded brown mushroom individuals. The improved algorithm is trained and tested on a self-built industrialized brown mushroom dataset. Compared to the original YOLOv7, the model's detection speed has increased by 15.5

The burgeoning of e-commerce and online platforms has led to an explosion in data volume and diversity of user preferences, making effective recommendation systems crucial for personalizing user experiences. While collaborative filtering algorithms are traditionally favoured for their ability to leverage user-item interactions, they grapple with data sparsity and noise challenges. To tackle these challenges, Various approaches have emerged in recent years to tackle these challenges. Recent strides in deep learning, particularly autoencoders and neural networks, have shown promise in addressing these issues. However, limitations persist, such as suboptimal feature extraction and the underutilization of combined nonlinear and linear latent features in traditional autoencoders, as well as the overlooked impact of active users in recommendations. Addressing these research gaps, this study introduces a novel recommendation algorithm that synergizes a deep denoising autoencoder with an attention mechanism, aiming to refine recommendation performance by mitigating data sparsity and enhancing feature extraction. This fusion approach innovatively combines nonlinear and linear latent features and incorporates a neural attention mechanism, significantly improving the precision and personalization of recommendations. Ultimately, the proposed algorithm's effectiveness is assessed and benchmarked against state-of-the-art approaches, demonstrating its potential to revolutionize recommendation systems by offering more accurate and user-tailored suggestions.

Pedestrian trajectory prediction plays an exceptionally vital role in autonomous driving, enabling advanced analysis and decision-making in certain scenarios to ensure driving safety. Predicting pedestrian trajectories is a highly complex task, encompassing static scenes, dynamic scenes, and subjective intent. To enhance the accuracy of pedestrian trajectory prediction, it is crucial to model these scenarios, extract relevant features, and fuse them effectively. However, existing methods only consider some of the scenarios mentioned above and extract static scene features through manual annotation of road key points, which fails to meet the demands of autonomous driving in complex traffic scenarios. To overcome these limitations, this paper introduces MINet – a network that employs multi-information feature fusion. Unlike previous approaches, MINet adopts a more automated approach to extract static scenes, including sidewalks and lawns. Moreover, the network incorporates pedestrian destination modeling to improve prediction accuracy. Furthermore, to tackle the challenge of collision avoidance in crowded spaces, this paper incorporates the extraction of dynamic scene changes through relative velocity modeling of objects. The proposed network achieved an improvement of 47.7

In recommender systems, the rating matrix is usually not a global low-rank but local low-rank. Constructing low-rank submatrices for matrix factorization can improve the accuracy of rating prediction. This paper proposes a novel network embedding-based local matrix factorization model, which can built more meaningful sub-matrices. To alleviate the sparsity of the rating matrix, the social data and the rating data are integrated into a heterogeneous information network, which contains multiple types of objects and relations. The network embedding algorithm extracts the node representations of users and items from the heterogeneous information network. According to the correlation of the node representations, the rating matrix is divided into different sub-matrices. Finally, the matrix factorization is performed on the sub-matrices for rating prediction. We test our network embedding-based method on two real-world public data sets (Yelp and Douban). Experimental results show that our method can obtain more accurate prediction ratings.