The load prediction of host resources is a key issue to enhance the cloud computing aid allocation system. With the change of cloud computing resource load displaying extra and extra complicated characteristics, traditional prediction algorithms can solely predict the linear traits of data, and it is tough to precisely predict useful resource usage. In order to enhance the forecasting accuracy of the model, a blended load forecasting algorithm based totally on machine learning is proposed. The machine learning prediction model can nicely match the nonlinear traits of the data. The linear phase of the algorithm makes use of ARIMA prediction, and the nonlinear section makes use of particle swarm optimization algorithm to optimize LSTM prediction. Then, the optimal least squares method is used to redistribute the prediction error weights of the autoregressive differential moving average model (ARIMA) and the long-term and short-term memory network models (LSTM), and finally the prediction results are output. The comparison experiment is carried out with the open actual load data set. The experimental results show that the prediction accuracy of the weight redistribution combination model is significantly higher than that of other traditional prediction models and machine learning prediction models when the prediction time efficiency is similar, and the real-time prediction error of resource load in cloud environment is significantly reduced.
As the world's most valuable cryptocurrency, Bitcoin offers a new opportunity for price forecasting because of its high volatility, which is much higher compared to traditional currencies. Since bitcoin prices fluctuate randomly over time, we can use a time series model to predict the price of bitcoin. For this purpose, we use the ARIMA model to predict the future bitcoin price based on past prices. The basic idea of the ARIMA model is that the data series of the predicted object over time is considered as a random series, and some mathematical model is used to approximate this series. Once this model is determined, it is possible to predict the future values from the past values of the time series as well as the present values. The model achieves high accuracy and robustness. The result shows that there's inevitable deviation every time the price trend is having acute change, and the deviation of actual value to predicted one is positively correlated to the average value.
This paper considers a new type of two-sided matching in which multiple numbers of agents are perfectly matched on both sides. Such matching can be used between multiple major students and laboratories. The many-to-many perfect matching problem cannot be solved by existing many-to-many matching algorithms, since the perfect property, which is a global property, cannot be represented by the participants’ preferences, which are local properties. This paper gives a DA(Deferred Acceptance) mechanism to match each student to the given number of different laboratories without a blocking pair by introducing a master list of students to resolve ties between students.
In recent years, the recommendation field has gradually started to combine GNN-like approaches to address the challenges. The Neural Graph Collaborative Filtering (NGCF) framework has made a preliminary attempt to extract structural knowledge in model-based collaborative filtering based on graph convolution with message passing mechanisms, opening up new research possibilities. However, the NGCF framework does not consider the semantic information in the topology and only constructs a single heterogeneous graph. In our work, we suggest explicit semantic encoding of edges for different user behaviors and propose a Heterogeneous Graph Convolution Collaborative Filtering (HGCCF) framework combined with message propagation mechanism, which can mine richer collaborative information and effectively alleviate the sparsity problem of bipartite graph and enhance the cold start capability. Furthermore, we reduce the computational effort through compressing the initial embedding vector and sharing parameters in the message passing. Our Top-N recommendation experiments on pre-processed real e-commerce data from Alibaba verify that HGCCF has higher recommendation accuracy and the ability to cope with cold starts. In addition, we also design hyperparametric experiments of HGCCF to explore the effect of HGCCF on performance with different propagation learning layers, different normalization coefficients prui, and different output dimensions of embedding propagation layers.
In recent years, event-based social networks have developed rapidly, and event recommendation has attracted more and more attention. At present, for event recommendation, it is centered on the event, and aims to help users get the events which they are interested in from a large number of events. However, compared with traditional recommendation problems, event recommendation has many challenges. First of all, there is no obvious explicit rating of users’ response to events, but implicit feedback. Secondly, the recommendation of events has heterogeneous social network relations. At the same time, most users participate in few events, which leads to a very serious data sparsity problem. In order to address these challenges and improve the effectiveness of event recommendation, this paper proposes an event recommendation model that integrates users’ online and offline heterogeneous social network information and time information. The model uses Bayesian personalized ranking as the framework to process the implicit feedback information of users and events, and simultaneously combines online and offline social network information and time information to model together to improve the accuracy of recommendation. Experimental results based on real data sets show that the performance of the proposed model is better than other methods.
Federated learning is a privacy preserving machine learning technology. Each participant can build the model without disclosing the underlying data, and only shares the weight update and gradient information of the model with the server. However, a lot of work shows that the attackers can easily obtain the client's contributions and the relevant privacy training data from the public shared gradient, so the gradient exchange is no longer safe. In order to ensure the security of Federated learning, in the differential privacy method, noise is added to the model update to obscure the contribution of the client, thereby resisting member reasoning attacks, preventing malicious clients from knowing other client information, and ensuring private output. This paper proposes a new differential privacy aggregation scheme, which adopts a more fine-grained hierarchy update strategy. For the first time, the f-differential privacy (f-DP) method is used for the privacy analysis of federated aggregation. Adding Gaussian noise disturbance model update in order to protect the privacy of the client level. We prove that the f-DP differential privacy method improves the previous privacy analysis by experiments. It accurately captures the loss of privacy at every communication round in federal training, and overcome the problem of ensuring privacy at the cost of reducing model utility in most previous work. At the same time, it provides a federal model updating scheme with wider applicability and better utility. When enough users participate in federated learning, the client-level privacy guarantee is achieved while minimizing model loss.
In this paper, the research and experimental analysis of cross-project application software defect prediction is carried out, and the TCA model is used to improve the application function of its prediction. The models pointed out in this paper usually include: normalization processing model and mathematical linear kernel mathematical statistics The difference between the functional SVM classifier and the extended migration component analysis TCA+ model is that the model pointed out in this paper not only satisfies the prediction of software defects within the project suitable for TCA, but also meets the prediction of software defects in the cross-project of TCA+, so the most appropriate normalization can be selected. Optimized processing options to improve cross-project software defect prediction capabilities.
Pre-trained language models such as BERT have proven essential in natural language processing(NLP). However, their huge number of parameters and training cost make them very limited in practical deployment. To overcome BERT’s lack of computing resources, we propose a BERT compression method by applying decoupled knowledge distillation and representation learning, compressing the large model(teacher) into a lightweight network(student). Decoupled knowledge distillation divides the classical distillation loss into target related knowledge distillation(TRKD) and non-target related knowledge distillation(NRKD). Representation learning pools the Transformer output of each two layers, and the student network learns the intermediate features of the teacher network. It has better results on tasks of Sentiment Classification and Paraphrase Similarity Matching, retaining 98.9% performance of the large model.
News recommendation aiming to find attractive news for users has been received many attentions in recent years. Existing news recommendation methods mainly focus on modeling user preference based on the interaction behaviors between users and news without the consideration of emotion information in the interaction. However, emotion information also plays an important role in improving the accuracy of news recommendation. In this paper, we propose an emotion analysis method for news recommendation with using multi-views to explore the impact of emotion information during the process of user's decision making. The emotion features extracted by the method are combined with the content features of the news to provide a comprehensive feature representation of the candidate news to improve the performance of recommendation. Experiments on real-world datasets show the effectiveness of the proposed method in improving accuracy of news recommendation.
Predicting stock prices and market indices is very difficult, and the associated prices and indices have too much uncertainty. There are already many deep neural networks for stock price prediction, which predict future stock prices based on historical stock price data. In this paper, a GRU-XGBoost model with attention is proposed to deal with heterogeneous data with various information in stock price prediction. The GRU model is used to solve the gradient problem, and the attention mechanism and XGBoost are used to save the context and process local optimal solutions. question. The experimental results show that the proposed method has better RMSE evaluation results.
Convolutional Neural Networks (CNNs) are popular in deep learning, especially in computer vision. But CNNs are poor at extracting spatial information, such as position and direction of entities in an image. A recently proposed model called Capsule network could effectively learn the spatial relations among entities, but it has weak feature extraction ability. The inception structure could extract multi-scale features. So this paper modifies the inception-v1 structure and adds it into Capsule network to strengthen the feature extraction ability. In the modified inception-v1, the max pooling branch is removed to reduce the loss of feature information. And Batch Normalization (BN) layer is added after each convolution to accelerate convergence and reduce over fitting. To organize categories information, a fully connected layer is added after digit Capsules. This paper conducted experiment on two public datasets, and the results show that the proposed model outperforms the original model in accuracy.
Wetlands are very complex yet pivotal ecosystems on Earth. They serve as habitats for various flora and fauna. Alongside, wetlands are crucial for biogeochemical exchange between the Earth’s surface and its atmosphere. A large proportion of organic carbon is sequestered in wetlands and plays a substantial role in the carbon cycle. The planning and management of wetlands depend a lot upon a reliable wetland model. The underlying complex dynamics of wetlands hinder the modelling of wetland extent. This study for the first time considers multivariate nonlinear dynamical system modelling using Nonlinear Autoregressive with Exogenous Inputs (NARX) model class. The data consists of weather variables and wetland fractions for two wetland sites falling under Asia and Africa. The model is simulated using fresh testing data and can predict wetland extent satisfactorily for both sample sites. The accuracy of the models is quantified using Root Mean square Error (RMSE) and Mean Absolute Error (MAE). A transparent NARX structure reveals the dynamical elements for the potential planning and management of wetlands.
For the optimization problem of deep learning, it is important to formulate a optimization method that can improve the convergence rate without sacrificing generalization ability. This paper proposes a layer-wise based Adabelief optimization algorithm to solve the deep learning optimization problems more efficiently. In the proposed algorithm, each layer of the deep neural network is set different learning rate appropriately in order to achieve a faster convergence rate. We also give the theorems that can guarantee the convergence property of Layer-wised AdaBelief method. Finally, we evaluate the effectiveness and efficiency of the proposed algorithm on experimental examples. Experimental results show that the converges speed of the layer-wised AdaBelief algorithm is the fastest compared with the mainstream algorithms. Besides, the new algorithm also maintaining an excellent convergence result in all numerical examples.
Neural vocoders are widely utilized in modern text-to-speech (TTS) and voice conversion (VC) systems due to their high generation quality and fast inference speed. Recently, GAN-based neural vocoders have attracted great interest due to their lightweight and parallel structures, which enable them to generate a high-fidelity waveform in a real-time manner. Most existing GAN-based vocoders adopt the Least Square GAN (LSGAN) training framework. In this paper, we analyze the weaknesses of the LSGAN waveform synthesis framework and, inspired by Relativistic GAN, propose a simple yet effective variant of the LSGAN framework, named Truncated Pointwise Relativistic LSGAN (T-PRLSGAN). In this method, we consider the pointwise truism score distribution of real and fake wave segments and combine the Mean Squared error (MSE) loss with the proposed truncated pointwise relative discrepancy loss to increase the difficulty of the generator to fool the discriminator, leading to improved audio generation quality and stability. To demonstrate the effectiveness and generalization ability of our method, subjective and objective experiments have been conducted based on Avocodo, UnivNet, HiFiGAN, ParallelWaveGAN, and MelGAN vocoders, which show a consistent performance boost over those typical LSGAN-based vocoders. Moreover, our T-PRLSGAN can support multiple types of discriminators, i.e. multi-scale wave discriminator(MSD), multi-period discriminator(MPD), and multi-resolution spectrogram discriminator(MRD), without modifying their architecture or inference speed.
Abstract: Intelligent communication network is the technical development trend of 5G and post-5G era. Big data analysis is the foundation of intelligent network while data isolation and privacy protection in network data analysis is a bottleneck problem. Federated learning is an emerging distributed machine learning framework which can make use of all parties' data for joint modeling while protecting the data privacy. In this paper, we propose an intelligent communication network framework which combines 5G Network Data Analysis Function (NWDAF) and federated learning to solve the above problem. Our work demonstrates that federated learning technology can ensure the data usage compliance in the process of 5G network intellectualization while solving the data isolation and data privacy protection problem.
The emergence of threat intelligence provides more foundation for tracing the source of network attacks, but it also necessitates a significant amount of manual analysis. Although data-driven automatic information extraction can effectively reduce labor consumption, it is limited by a lack of labeled data in the field of threat intelligence. To overcome this limitation, we propose TRAPPER, a threat entity recognition framework that can infer real threat entities from unlabeled threat sentences, avoiding the difficult labeling work. TRAPPER relies on label functions and three components, label aggregator, label predictor, and label expander, which guides the model with weak supervision and uses transfer knowledge as an aid. The label functions permit us to inject expert knowledge into the label aggregator to generate the inputs needed by the label predictor. It enables the label predictor to learn to recognize threat entities. The label expander combines the multi-source noisy label information with the transferred entity recognition semantic knowledge to further expand the entities. Throughout the process, the components promote each other by learning from each other. Comparative experiments on three threat intelligence-related datasets show that our method can effectively identify threat entities and achieve a maximum F1 score improvement of 6.3% over the best baseline.
The rapid growth of large-scale trajectory data has prompted researchers to develop multiple large-scale trajectory data management systems. One of the fundamental requirements of all these systems, regardless of their architecture, is to partition data efficiently between machines. In the typical query operations of tracks, the query on ID is a frequent operation of track query, such as ID time range query, ID space range query, etc. A widely used ID indexing technique is to reuse an existing search tree, such as a Kd-tree, by building a temporary tree for the input samples and using its leaf nodes as partition boundaries. However, we show in this paper that this approach has significant limitations. To overcome these limitations, we propose a new index, BOD-tree, which inherits the main features of the Kd-tree and can also partition the dataset into multiple balanced splits. We test the method on real datasets, and extensive experiments show that our algorithm can improve resource usage efficiency. \end{abstract}
Activation functions are integral components in neural networks because they are the calculations that each neuron performs on their inputted data before outputting it to the next neuron to help the neural network match the ground truth of the data sooner, thus converging faster. However, popular activation functions are not parameterized and those that are, have too few parameters, therefore lacking the ability to fully train the shape of the activation function. This paper introduces RPSigmoid, an activation function based on the Sigmoid function, and with four additional parameters which represent the vertical stretch factor, horizontal stretch factor, angularity, and slope of the asymptotes (which might be horizontal or oblique) of the sigmoidal curve. These parameters are randomized within a range before training and their values are updated along with other neural network parameters during backpropagation. Affirmative results of RPSigmoid present neural network training with a low-resource approach to yield impressive training results.
Open resolvers can be easily exploited by malicious actors to launch DDoS attacks against important services on the Internet, which has aroused much concern in the Internet community. Researchers have studied extensively the population, structure, and malicious behavior of open resolvers, while little has been done to reveal how open resolvers respond to non-recursion queries. In this paper, we conduct an Internet-wide measurement on the recursive behavior of open resolvers. We discover that more than 1 million gushing resolvers in the wild are more enthusiastic than needed to respond to non-recursive queries, either triggering a new recursive resolution process or replying with cached records. Furthermore, we discuss possible security implications posed by the massive gushing resolvers. Specifically, we show that gushing resolvers are prone to be targets of ranking manipulation attacks if they happen to be the data collection points of top lists. At last, we put forward suggestions for resolver operators to improve such a situation.
Abstract: in the development of information environment, rural areas of e-commerce to achieve urban and rural development of integration, information, supply chain logistics needs to develop. In the era of e- commerce development, supply chain management makes the speed and scope of information sharing, the efficiency of information processing, the mass storage of information, the value of information has been significantly enhanced, especially in rural areas, the circulation system is less developed, it is more necessary to study the local circulation supply chain system and put forward feasible optimization measures. At the same time, supply chain management, as one of the integrated management methods, has been paid more and more attention, especially in the era of e-commerce development, the research on the construction of rural circulation supply chain system is not only for the realization of rural circulation modernization, but also for the promotion of rural economic development, it is of great significance to solve the problems of agriculture, rural areas and farmers in our country, and it also provides an important theoretical basis for the development of rural circulation. Based on the study of rural e-commerce ecological chain and value chain, this paper constructs the rural e-commerce supply chain system, it is of theoretical and practical value to analyze how the rural e-commerce supply chain system with e-commerce service center works more effectively.
