In the rapidly evolving landscape of e-commerce, online shopping has gradually gained widespread acceptance. Yet, physical retail channels continue to exert a significant influence on product sales. Within this framework, sales forecasting occupies a pivotal position in traditional commerce and is indispensable for guiding corporate strategy formulation. However, most existing forecasting models fail to fully harness the potential impact of multi-source external information on consumer purchasing behavior. To address these challenges, this study proposes a novel sales forecasting model that synergizes the Extreme Gradient Boosting (XGBoost) algorithm with multi-source data integration techniques. Not only have we comprehensively collected an extensive range of external data, but we have also integrated the embedding and aggregation techniques of Point of Interest (POI) data to enrich the set of external information features available to the model. Leveraging feature engineering techniques, these heterogeneous data are transformed into formats amenable to model training. Relying on the nonlinear modeling capabilities of XGBoost and its efficacy in handling large-scale datasets, we trained and optimized the model. Empirical results indicate that our proposed method, predicated on multi-source data integration, significantly outperforms traditional models based on a single data source, thereby enhancing prediction accuracy and providing more precise inventory management strategy support for businesses.
Objective: Because of the aging of parts and mechanical structure, the vibration of connecting rod is easy to produce abnormal state, while the abnormal vibration states are difficult to detect in a timely manner,thus resulting in shutdown and material breakage during the tobacco processing process, or uneven tobacco screening, thereby reducing the quality of tobacco products . In order to solve this problem, an automatic detection technology based on Dempster-Shafter (DS) theory is proposed in this paper, so as to deal with the abnormal vibration of connecting rod in time. Methods: The image data inside the tobacco sieve vibrating groove were collected, and the Kanade-Lucas-Tomasi (KLT) algorithm was used to track the continuous frames of the image and extract the connecting rod feature points at adjacent moments of the tobacco sieve vibrating groove. A series of microphones are constructed to obtain the location of noise source signals, and the noise source signals are denoised using wavelet coefficients. DS theory was used to fuse the image features and sound signals, calculate the vibration state conflict factor of connecting rod, judge the vibration state of connecting rod according to the conflict factor, and complete the automatic detection of abnormal vibration of connecting rod in the vibration groove of tobacco screen. Results: The experimental results show that the proposed method can accurately identify the fault position of the connecting rod and classify the abnormal vibration level, and the results of abnormal vibration amplitude of the connecting rod in the vibrating groove of tobacco screening are highly consistent with the actual results. Conclusion: The vibration anomaly automated detection technology designed in this article can be used for the analysis of the state of connecting rod mechanical components in the tobacco industry and the diagnosis of faults.
Breast cancer remains a major global health concern, and early diagnosis is crucial for improving patient outcomes. This study explores the performance of Fourier Convolutional Neural Networks (FCNNs) in comparison to traditional Convolutional Neural Networks (CNNs) like VGG16 for classifying breast cancer using ultrasound images. We evaluated the models on key performance metrics, including accuracy, precision, recall, and AUC-ROC, while also considering computational factors such as training time and memory consumption. Results show that VGG16 consistently delivers stable performance across different learning rates, while FCNN16 significantly outperforms VGG16 at a lower learning rate of 0.0001, achieving near-perfect classification metrics. These findings highlight the potential of FCNNs for breast cancer diagnosis, though further research is necessary to fully optimize this approach.
Histopathological image analysis remains at the forefront of computational pathology presenting numerous challenges and demanding tasks, primarily due to the complex nature of tissue structures and the extensive scale of whole slide images (WSIs). Deep learning models have been widely used in histopathology image analysis, especially convolutional neural network (CNN)-based models for classification. However, CNNs have certain limitations due to their small receptive field. Recent works employed adaptations of the classical transformer architecture to visual data [1] [2]. Models such as Vision Transformer (ViT) and Swin Transformer leverage the powerful multi-head self-attention mechanism and have demonstrated comparable or superior performance to state-of-the-art CNN-based classification models. Despite their successes, these models require huge amounts of training data to effectively learn representations as they lack the inherent inductive biases of CNNs. This work compares Vision Transformers with baseline CNN models using a breast cancer histopathological dataset. Further, we employ a novel knowledge-distillation approach to enhance the learning efficiency of vits, When trained with a limited amount of data, Unlike previous works, we aimed to minimize convolution operations when generating patch embeddings to preserve spatial information before reaching the transformer attention layers, we achieved an accuracy of 87.7% for the ViT-base trained as a student of ResNet50, which represents a 1.2% improvement in accuracy over the standalone ViT-base. [3]
Considering the uncertainty of stock data, this paper studies interval-valued fuzzy portfolio decision model with ESG and some realistic constraints. First, a screening process is employed to identify investable stocks based on their ESG ratings. Then, a multi-period portfolio decision model with short-selling constraint is established by maximizing terminal wealth and minimizing terminal risk. By using the weighted programming method the proposed portfolio model is transformed into a single-objective model to solve the optimal decision strategy. Finally, a numerical example of stock data from the Shenzhen Stock Exchange is given to illustrate the efficiency of the presented portfolio decision model.
Interval-value Fermatean hesitant fuzzy sets (IVFHFSs) are new data model for dealing with complex, uncertain information. However, the researches on IVFHFSs score functions (SFs) have low discrimination rate and existing multi-attribute group decision-making under IVFHFSs are scarce and have no ability to classify. Therefore, this paper establishes a new three-way multi-attribute group decision-making (3W-MAGDM) based on IVFHFSs. First, novel IVFHFSs SFs are proposed. Next, the objective conditional probabilities calculation method is derived by using the probabilistic dominance relation and the relative loss functions calculation method in three-way decision is developed. The subjectivity of 3W-MAGDM under IVFHFSs is greatly reduced. Finally, a new 3W-MAGDM framework based on IVFHFSs is constructed. The new approach has a high discrimination rate in SFs and it not only has a ranking function but also has a categorization function.
Acoustic analysis is an objective measurement approach commonly used in the assessment of the singing voice. It involves parameters such as fundamental frequency, formant, intensity, vibrato pattern, and others, providing parameters on the internal view of the singing voice. However, presenting multi-dimensional voice data from multiple audio files simultaneously in current studies is a challenging task. This review summarises existing methodologies, highlighting their limitations and proposing future directions, particularly integrating Information and Communications Technology to enhance vocal analysis visualisation. This study also emphasises the critical role of advanced visualisation methods in enhancing the accuracy, accessibility, and impact of vocal analysis, and it aims to enhance the understanding of vocal mechanisms, support singers and students in vocal practice, and drive forward the field of vocal research through innovative, user-friendly technologies. Our future work will focus on developing interactive visualisation techniques to simplify and enrich the interpretation of vocal data, making complex acoustic properties accessible to a broader audience, including non-experts.
Correlation analysis has been a powerful paradigm to discover and analyze hidden properties and patterns of large-scale datasets for decades. At now, correlation analysis turns to be a perfect tool for supporting big multidimensional data analysis and mining, with a wide range of relevant properties, including the amenity of supporting meaningfully exploration and discovery of multidimensional ranges kept in such kind of datasets. These operators are thus the basis for several multidimensional big data analytical tools that can be designed and implemented on top of the foundations defined by correlation functions. In line this this scientific area, the talk will provide introduction and motivations, models and algorithms, and, finally, best-practices guidelines for effective and efficient implementations of correlation-analysis-based tools over big multidimensional datasets.
Power system load forecasting is the foundation of power system operation and planning. Meteorological factors such as temperature, humidity, and wind speed have a significant impact on the load of the power system. In recent years, with the popularity of data-driven methods, power system load forecasting has developed. However, in existing research on power systems, the study of relationships between variables is predominantly based on correlation, with little mention of the concept of causality. In this paper, an adaptive time window strategy is proposed to identify time periods where causal relationships may change. Then, the Time Causal Discovery Framework (TCDF) was used at each time window to explore the causal relationship between electricity load and climate data. Finally, compared with the traditional correlation-based analysis method, the experiment shows that our method not only explains the causal relationship between power load and climate, but also effectively improves the prediction performance.
In the fields of computer vision and intelligent recognition, with the increasing complexity and dynamism of crane scenes, how to use digital twins technology and artificial intelligence technology to improve the accuracy and efficiency of crane image recognition has become a research hotspot. Based on the combination of digital twin technology, convolutional neural networks, and long short-term memory networks (DTE-CNN-LSTM), an algorithm has been developed to achieve intelligent recognition and semantic understanding of complex crane operation scenes by introducing a virtual simulation environment for image feature optimization and processing. The research results indicate that the DTE-CNN-LSTM algorithm performs well in multiple crane scenarios. After optimization, the recognition accuracy of night cranes reached 100%, the recognition accuracy of rainy scenes increased to 95%, and the misclassification rate decreased to 5%. The recognition accuracy of obstacle has reached 98%, and the recognition accuracy of crane operation area has reached 90%, significantly reducing the misclassification rate. Overall, the DTE-CNN-LSTM algorithm achieved an accuracy rate of 96.70% and an F1 value of 96.64% in classification tasks for different crane scenarios. Its robustness and generalization ability in complex crane working environments have been verified, demonstrating high potential for application.
Satellite image transmission is crucial for distant sea monitoring and navigation safety. Beidou satellite communication has low cost, high security and confidentiality, and has broad development potential in the field of satellite image transmission. However, the communication capacity of Beidou satellites is limited. In order to increase the transmission rate of images, the images are compressed and sent. The shore end receives and decodes images with lower resolution. Therefore, the shore end designs super-resolution reconstruction technology to perform image processing on the received images. The improvement of resolution, improving the quality and texture details of images and making them clearer and more realistic are urgent problems that need to be solved. This paper proposes a generative adversarial network super-resolution reconstruction method based on an adaptive mechanism. This method uses a new perceptual loss function combination to integrate pixel loss, feature loss based on an adaptive weight mechanism, and adversarial loss to achieve a total loss function. At the same time, an adaptive weight mechanism is introduced in the feature loss, and a spatial attention mechanism is introduced in the generator and discriminator of SRGAN, thereby effectively distinguishing the salient areas and background areas of the image, and fully retaining the high-frequency details and edge features of the image. This solves the problems of low resolution, edge smoothness and detail distortion of images received at the shore end. Experimental results show that the image details of the overall area of the ship image (including the salient area and the background area) reconstructed by this method are clearer than existing super-resolution reconstruction methods such as EnhanceNet, ESPCN, and SRCNN, and the edge integrity of the salient area is better.
The motivation behind fuzzy logic in data mining is to address the inherent uncertainty and imprecision in real-world data and make the mined results more interpretable for humans. Temporal Fuzzy High Utility Itemset Mining, which incorporates transaction time, is an emerging field with significant potential for analyzing time-sensitive data. Although several studies in this area have been conducted, for instance, recently fuzzy list-based approaches, a significant challenge remains in joining operations of conditional fuzzy lists when generating candidate itemsets. To solve this, we have proposed a pruning strategy based on item co-occurrences to reduce the number of join operations using anti-monotonic property. Experiments on real datasets show our approach outperforms traditional algorithms in terms of runtime and candidate generations with little memory overhead, up to 95% of non-promising candidates are pruned.
This paper addresses the challenges of managing and processing unstructured or semi-structured text, particularly in the context of increasing data volumes that traditional linguistic databases and algorithms struggle to handle in real-time scenarios. While humans can easily navigate linguistic complexities, computational systems face significant difficulties due to algorithmic limitations and the shortcomings of Large Language Models (LLMs). These challenges often result in issues such as a lack of standardized formats, malformed expressions, semantic and lexical ambiguities, hallucinations, and failures to produce outputs aligned with the intricate meaning layers present in human language. As for the automatic analysis of linguistic data, is well known that Natural Language Processing (NLP) uses two different approaches, coming from diverse cultural and experiential backgrounds. The first approach is based on probabilistic computational statistics (PCS), which underpins most Machine Learning (ML), LLMs, and Artificial Intelligence (AI) techniques. The second approach is based, for each specific language, on the formalization of morpho-syntactic features and constraints used by humans in ordinary communication activities. At first glance, the second approach appears more effective in addressing linguistic phenomena such as polysemy and the formation of meaningful distributional sequences or, more precisely, acceptable and grammatical morpho-syntactic contexts. In this paper, we initiate a scientific discussion on the differences between these two approaches, aiming to shed light on their respective advantages and limitations.
Q-rung orthopair hesitant fuzzy set (q-ROHFS) is a powerful instrument for addressing uncertainty problems. Nevertheless, the classification methods of three-way multi-attribute group decision-making (TWD-MAGDM) under this new model have been seldom researched, and the current TWD-MAGDM method in a hesitant fuzzy environment fails to consider the psychological behavior and fuzzy correlation of decision-maker, resulting in not enough distinction among classified objects. To resolve this issue, we present a novel TWD-MAGDM classification model for a q-rung orthopair hesitant fuzzy (q-ROHF) environment. Firstly, this paper considers the fuzzy correlation by allocating weight through Shapely and combines the prospect theory and Gaussian function to develop a preference function that can accurately describe the loss and gain. Based on this function, it presents a relative utility function that can more accurately measure the utility. Secondly, we provide a conditional probability that considers psychological factors and has enhanced recognition capabilities. Finally, a novel TWD-MAGDM classification model for q-ROHF is provided based on the new relative utility function and conditional probabilities. We subsequently verify the efficacy of the proposed approach.
Corrugated cardboard is a crucial component of packaging materials, and its surface defects directly affect the quality and aesthetics of the products. To address the challenges posed by the diverse and uneven distribution of defects on cardboard surfaces, we propose the YOLOv8-GSP, a detection algorithm based on attention mechanisms and lightweight improvements in YOLOv8 for surface defects of corrugated cardboard. Firstly, to tackle the insufficient size of the cardboard defect dataset, Projected GAN was employed to augment the dataset, thereby enhancing the training effectiveness of the detection model. Then, Ghost convolution was introduced to reduce the computational complexity and improve the operational speed. Additionally, an SE attention module was incorporated into the Neck network to emphasize the feature information of the corrugated cardboard surface, thereby enriching feature fusion. We improved YOLOv8 by incorporating Ghost convolution and attention mechanisms to enhance the efficiency and accuracy of the detection model. These improvements not only reduce the computational complexity of the model but also enhance detection performance by better focusing on target features. Our method is experimentally validated on a self-constructed dataset, providing a more efficient and accurate solution for the automatic detection of corrugated cardboard surface defects. Indeed, the proposed method’s mean Average Precision (mAP) increased by 1.6% compared to the standard YOLOv8.
Sequential recommendation aims to predict users’ next preferred items according to their interaction sequences. Existing methods mainly utilize user-item interaction information, which may suffer from the issue of semantic information loss. In the paper, a Meta-Path guided Pre-training method for sequential Recommendation (MPPRec) is proposed to capture rich and meaningful semantic information between users and items. Specifically, MPPRec firstly learns the node embeddings guided by meta-paths in the pre-training phase. Then, the node embeddings are optimized according to task in the fine-tuning phase. Extensive experiments conducted on four real datasets demonstrate MPPRec outperforms the baseline methods.
The integration of Generative Artificial Intelligence (GenAI) with the Metaverse for a next-generation education is a complex but challenging task. The GenAI-enhanced Metaverse classrooms require innovative instructional designs that use virtual reality and augmented reality to enhance engagement and personalized learning. Educators must adapt to new roles over traditional teaching methods, while learners need to develop digital literacy skills that are essential for navigating and inhabiting in these environments. Such learning environments require significant advancements in real-time processing, scalability and interoperability of different platforms, while ensuring data privacy and security. The equity of access to high-speed internet and advanced devices still remains a serious barrier, which can increase the potential existing inequalities between different educational environments. Ethical considerations, including the responsible use of GenAI, the creation of unbiased educational content, and the psychological impacts of extended usage of virtual reality, are also of important consideration. The aim of the paper is to explore in detail the different challenges through a comprehensive analysis of the obstacles and potential solutions and to propose a collaborative framework involving educators, technologists, policymakers and industry stakeholders to address the effective implementation of the integration of GenAI and the Metaverse for a next generation education.
In recent years, the education landscape has undergone major changes, such as the pandemic, the introduction of digitalization, and advanced technologies that facilitate the dissemination of contactless services. All of these are driving innovations in the learning experience. Furthermore, the United Nations has declared a digital bill of rights, which seeks to define digital as a fundamental human right beyond the concept of expanding digital access or bridging the gap. However, the issue of poor information accessibility for the disadvantaged remains a significant challenge, both domestically and internationally. In order to face this reality, this paper implemented a function based on LaTex that allows the visually-impaired students to enter formulas by voice in there reports. The performance and usability evaluation of this feature was analysed by T-Test, and it was discovered that the report generator has a significant effect on enabling independent learning for blind students. Therefore, this feature is expected to not only foster independence in learning for blind students in science and technology, but also to ensure general social integration and digital accessibility for blind students through the introduction of technology that is in line with the times.
Image processing has become a central topic in the era of big data, particularly within computer vision, due to the growing volume and diverse resolutions of images. Low-resolution images introduce uncertainty, underscoring the need for high-performance classification methods. Convolutional Neural Networks (CNN), especially the U-Net architecture, are widely applied for pixel-level segmentation due to their encoder-decoder structure. This study applied U-Net on a CT scan image dataset to segment lung images, followed by a CNN classifier to classify lung cancer stages (I, II, IIIa, IIIb). The U-Net model outperformed standard CNNs, achieving 99% in accuracy, precision, sensitivity, and F1 score, compared to the conventional CNN’s 97%, 95%, 97%, and 96%, respectively.
In this study, we investigate the potential of the Generalized Sigmoid (GS), σ(x;a,b) = (1 + e–a(x – b))–1 as an adaptive activation function within the YOLOv8 framework for drug detection and classification tasks. Medication errors, particularly in identifying pills and tablets, present serious challenges in healthcare due to visual similarities and environmental factors. We propose the use of the GS, which allows flexible parameter tuning, to enhance detection accuracy. Various configurations of the GS are evaluated against the widely-used SiLU and standard sigmoid functions. Our experiments reveal that optimal tuning of GS parameters, specifically with a = 1.50 and b = 1.50, improves model performance in terms of mean Average Precision (mAP) and overall classification accuracy. These findings suggest that the GS can surpass standard activation functions, offering a promising advancement in deep learning applications for healthcare.
