With the continuous increase of online courses and the rapid growth of network data, how to improve the recommendation accuracy and real-time performance of personalized recommendation system is a key issue. In order to improve the recommendation quality and real-time performance of the recommendation system, this paper uses SSM (Spring, Spring MVC, Mybatis) Framework, the most popular and latest framework of the enterprise, as the framework of the recommendation system, and uses the hybrid recommendation algorithm (Socialized Recommendation algorithm) that integrates Collaborative Filtering recommendation algorithm and social relationship to make course recommendation. In this paper, the course and relational data of scholar.com are used as data sets for research, so as to provide richer social relationships and more references for experiments.
With the continuous development of deep learning, convolutional neural networks are gradually playing an increasingly important role in classification models. Adding attention mechanisms based on convolutional neural networks to improve the classification effect of the model has been proven a lot. However, only theoretically understanding how attention mechanisms enhance the model's performance is a lack of intuition. Therefore, to explore the role of the attention mechanism in the model, this article will start from visualization and analyze how the attention mechanism improves the model's performance in combination with theory. The method makes attention more intuitive and more accessible for non-professionals to understand the working principle of attention mechanism, facilitating the promotion of attention mechanism and convolutional neural networks.
For catching and processing seasonality in seasonal time series, a combining forecasting method based on seasonal unit root test, i.e., the Dickey-Hasza-Fuller (DHF) test and support vector regression (SVR) is proposed, which is denoted as DHF-SVR method. The DHF-SVR method employs DHF test to identify seasonality in series and utilizes seasonal differencing operator to process the seasonal time series; for solving the difficulty of adaptive selection of maximum lag order, a SVR hyper-parameters tuning method based on a genetic algorithm (GA) with real-integer hybrid encoding is proposed. The experimental comparison demonstrates that the proposed DHF-SVR method could improve the forecasting performance in comparison with the comparative methods.
In this paper, a resource allocation algorithm based on deep reinforcement learning (DRL) is proposed to solve the problem of co-existing enhanced mobile broadband (eMBB) slices and ultrareliable low latency communication (URLLC) slices based on Non-orthogonal Multiple Access (NOMA) in downlink network scenarios. Double Deep Q network (DDQN) is designed to output subcarrier and power allocation simultaneously. In addition, expert data is added in the training process to accelerate network convergence, and our goal is to optimize the spectral efficiency of the system. Simulation results show that compared with the baseline based on heuristic joint user pairing and power allocation algorithm [1] and Orthogonal Multiple Access (OMA), proposed algorithm can achieve higher spectral efficiency and ensure isolation between slices.
In recent years, the development of green ports has attracted great attention from the academic circles, port management departments and port and shipping enterprises. In the past in the construction of port development, we have too much emphasis on port throughput and other economic indicators, although to a certain extent, contributed to the rapid development of trade, but due to global climate change and energy consumption problems intensified, the concept of green development widespread attention from the society from all walks of life, has been given to the mission of green port is imminent. At present, there are not enough literatures about green ports, and the few literatures remaining are mostly focused on a single dimension, which cannot deeply analyze the information about green ports. This paper attempts to explore green ports through innovative research and analysis, hoping to be helpful to the research on green ports.
With the increase of people’s work activities indoors, indoor positioning technology has become a hot spot in the field of positioning technology. The current mainstream indoor positioning includes WiFi, infrared, Bluetooth, ultra-wideband, ZigBee, RFID and ultrasonic technologies, each of which has its own advantages, but there are also certain shortcomings. The paper is based on WiFi localization technology, incorporating Pedestrian Dead Reckoning (PDR) localization technique and extending Kalman filter algorithm to solve the fused data. The experiments prove that the positioning error of the fused WiFi and PDR indoor positioning methods is smaller than that of the two technologies alone. The maximum error of the combined positioning method is 1.9532 m, the minimum error is 0.4727 m, and the mean error value is 0.8491 m. Compared with the two separate positioning methods, the accuracy of the Kalman filtered indoor positioning method fusing WiFi and PDR improved by 42.57% relative to PDR accuracy and 31.1% relative to WiFi accuracy, thus verifying the effectiveness of the Kalman filtered indoor positioning accuracy improvement by fusing WiFi and PDR.
The ICP algorithm has been widely used for point cloud registration, but it has low computational efficiency under large-scale point cloud data. Moreover, it only uses the structural information of point cloud in the solving process. Its accuracy and efficiency will be greatly affected in scenes lacking of obvious structural features and scenes with low overlap. In this paper, we propose a semi-dense ICP algorithm based on the neighborhood of SIFT feature points. Instead of the whole points or the sparse SIFT feature points, it selects the neighborhood of SIFT feature points as the matching range of the nearest matched point pair. We also introduce a new objective function with adaptive weights for the ICP algorithm, which balances the importance of the structural features and image features by dynamically adjusting the weights, and ensures that the ICP iteration can converge correctly. Experimental results show that the proposed method achieves higher accuracy and efficiency than several related methods. Especially, its effectiveness is also verified in scenes without obvious structural features or texture features and scenes with low overlap rate of two frame point clouds.
The recommendation system achieves user preferences by analyzing users’ historical behaviors, find user groups similar to target users, predict target users’ ratings of items of interest and make recommendations. Therefore, various techniques have been proposed to develop similarity measures. Research on recommendation algorithms that integrate user trust information has made considerable progress. However, only the definition of trust relationship based on the degree of social relationship can’t truly reflect the trust relationship between users. Focusing on the shortcomings of current trust algorithms, this paper proposes an implicit trust calculation method that integrates user entropy. By using the information entropy of user ratings to improve the previous similarity measurement, the user’s trust degree is integrated to make more accurate selection of neighborhoods, and the situation of high trust and low interest similarity is reduced. The effectiveness of the algorithm is verified through experiments, it is proved that it is superior to the previous similarity measure, and compared with the traditional recommendation algorithm, the prediction accuracy has been improved.
Aiming at the problem of the unbalanced advertising user data of social networks leading to unsatisfactory prediction results, we propose a prediction model for advertising users based on the combination among K-Means, synthetic minority oversampling Technique (SMOTE), and Ensemble Learning. On the basis of the real user data provided by Scholat, we analyzed the data and extracted many key features from it to draw a portrait of advertising users. Our algorithm first clusters the minority class, and then processes the continuous and discrete features of each sample separately through the improved SMOTE to synthesize new minority samples, and finally constructs an integrated classifier using the ensemble learning. This method effectively avoids the problems of blurred positive and negative class boundaries caused by SMOTE and the inability of SMOTE to process discrete features. Meanwhile, ensemble learning enables the classifier to get more reasonable results and reduce overall errors. The experimental results show that our method improves the quality of the generated minority class samples and significantly improves the prediction performance of advertising users.
Accurately diagnosing the cause of the failure type of the automatic sorter equipment is the basis for accelerating the efficiency of logistics sorting operations. Based on the research on the diagnosis problem of the parcel sorter equipment, based on the analysis of the fault text data, a method of parcel sorter equipment fault diagnosis based on association rules is proposed. First, build a two-layer fault diagnosis model based on the fault text information and expert experience; use the TF-IDF method to extract the semantic features of the fault text, and propose an improved Apriori algorithm on the basis of the traditional Apriori algorithm to mine the fault text information. The law of association between. The research results show that the evaluation indexes of the improved Apriori algorithm are higher than those of the Apriori algorithm, which proves the feasibility of the method.
In some emergency scenarios, UAV relay network is often used to provide temporary communications services. Different from the absolute position information, this paper proposes a deployment scheme of UAV network which is based on the relative position information. As a dynamic scheme, it contains several strategies for the deployment, retrieval, and topology control. Compared with other two schemes based on the absolute position information, greedy-based scheme and genetic algorithm-based scheme, the network based on our scheme achieves 90% and 70% of the coverage performance of them respectively and significantly outperforms those two static schemes in robustness and load balancing.
Since its appearance, neural network and time series related works had been tremendously developed. The GNN, CNN based Graph Convolutional Network was proposed recently, and made obvious contribution in solving link prediction and vector classification problems. With the development of chain operation, more and more data analysis institutions had started to provide network based sales and inventory level prediction service. This paper dedicated to propose a new prediction mean by joint employing the network analysis mean and single vector time series prediction mean to achieve a significantly more accurate prediction result. Employing GCN as the network analysis part and the ARIMA model as single vector time series prediction part. The GCN model is especially suitable for analyzing embedding information on non-Euclidean graph while the ARIMA model is able to utilize rather small amount of data to achieve an accurate prediction result, by combining these two models, this paper have achieved a better prediction result on non-Euclidean graph. And a supply chain related dataset was used to certify the efficiency of GCN-ARIMA model.
In recent years, with the deepening of power system architecture adjustment and market-oriented reform, smart microgrid has become the main development direction of power grid architecture. In order to meet the higher requirements of decentralization, autonomy, intelligence and marketization in the process of power grid reform, digital technologies such as artificial intelligence and block chain should be introduced as the key support. In view of the consistency between the technical characteristics of block chain and the development demand of power grid, a power trading system model based on alliance chain is proposed. In order to realize the power balance and maximize the benefits of the microgrid within the power grid, a trading strategy optimization scheme based on MADDPG algorithm is proposed. The experimental results show that the algorithm can help microgrids to formulate the trading strategy which is most in line with the overall benefits of the grid and maximize the total revenue of the system. The performance of this algorithm is better than DDPG algorithm and random trading method.
Science and technology resources can be regarded as web services on the Internet, in order to realize the reuse of science and technology resources, the industry provides services for Internet users in multiple web service methods. In order to reuse of the web services, multiple web services need to be combined according to certain rules and business logic to solve the problem of limited functions of a single web service. The web service composition algorithm focuses on finding a service composition scheme with the best service quality. A single service composition scheme cannot cope with the dynamic changes of the network environment in real time, such as service failures. This paper proposes a graph-based service composition method, which uses the service dependency graph to establish the relationship between services, and combines the functional and non-functional attributes of the service. In the service selection stage, a formula for calculating the matching degree between the service and the target parameters is proposed to evaluate the current matching degree between the service and the target parameters, and the service with the best matching degree is selected.
Finding the shortest path in a transportation network is an important and common problem in practice. A* is usually recognized as an efficient approach to the shortest path problem. However, the time and space requirements make A* hard efficiently find the shortest path in a massive and complicated transportation network. Reducing useless nodes generated is an efficient way to improve A*. In view of spatial distribution characters of transportation networks, this paper presents a fast shortest path searching method named A*_SA. The main idea of A*_SA is that when expanding node n, only those children whose h-cost is equal to or smaller than h(n) + C are generated and stored. Motivated by simulated annealing, in order to avoid getting stuck in local optima, the children with larger h-cost are also be stored with a certain probability. Experimental results prove the efficiency of A*_SA.
Gradient descent method is the preferred method to optimize neural networks and many other machine learning algorithms. Especially with the wide use of deep learning in recent years, gradient descent algorithm has become more and more important. In gradient descent algorithm, learning rate is a very important parameter. The setting of learning rate directly affects the performance of the final model. The existing learning rate optimization algorithms adjusts learning rate based on the idea of step-by-step reduction. Different from this idea, this paper based on human walking law proposes a new optimization algorithm, the consolidate step-by-step algorithm (CSBS), which determines the learning rate according to the gradient of each iteration. In this paper, MNIST data set is used to verify the performance of the algorithm. The experimental results show that the CSBS algorithm accelerates the convergence speed of the model and reduces the sensitivity to the initial parameters.
With the development of the Internet and 5G era, RCS (Rich Communication Suite) based on native SMS interface is developing at a high speed, which will develop a multi-functional service to provide users with picture, audio, video and other forms of interaction. In this context, frequent messaging between users or interaction between users and third-party enterprises will produce a large number of RCS logs. These logs play a very important role in the credit granting of electronic vouchers and the traceability of message source documents. This paper proposes a 5G message log credit management and verification system based on blockchain. Under the background of traditional 5G message service, combined with blockchain technology and distributed storage, it realizes the decentralization, trust and tamper resistant ability of RCS logs, so that users can trace and authenticate RCS logs. The paper also proves the feasibility of the scheme, and analyze its good retrieval efficiency and throughput requirements through experiments.
Task allocation is a key technology in the research of mobile crowdsensing. The previous research only focused on single-task allocation, and seldom considered the monopoly nature of tasks, quality requirements, and the constraint relationship between tasks. This paper comprehensively considers the above factors and designs a multi-task allocation scheme for mobile crowdsensing to maximize the profit of the service platform. First, divide the tasks into monopoly tasks and non-monopoly tasks, and judge whether they will be executed according to the profit that monopoly tasks can bring to the platform; For non-monopoly tasks, an efficient allocation plan is designed based on genetic algorithm and greedy algorithm; Secondly, considering the quality requirements of tasks and the constraint relationship between tasks, comparing the existing classic task allocation schemes, simulation experiments verify that the proposed algorithm has better effects in terms of platform profit and task coverage.
Pedestrian detection in the scene of automated container terminal can bring massive benefits to improving operation safety of the terminal. However, pedestrian detection in container terminal environment exists small pedestrian target size, extreme foregroung-background class imbalance problems. To solve these problems, this paper proposes a one-stage pedestrian detection model based on the improved ResNet. For small target pedestrian detection, we introduce the Channel attention module and spatial attention module to ResNet, enhancing the network’s attention to small target pedestrians. Moreover, the feature pyramid network is added to achieve semantic fusion of different scales to enhance the detection of small-scale targets. Finally, the focal loss function is employed as the loss function to suppress the weight influence of simple samples and relieve the problem of extreme foregroung-background class imbalance. The experimental results on the virtual container terminal pedestrian dataset show that the average accuracy of the model is 74.3%, which is 3.9% higher than that of ResNet. The detection accuracy is also improved by comparing with traditional pedestrian detection models. In addition, the effectiveness of the model is verified on Caltech and COCO2017 datasets.
Big data technology can play a significant role in exploring and analyzing classical Chinese literature and in enhancing our understanding and promotion of traditional culture. Analyzing psycholinguistic words used in ancient people’s self-expression texts is a good way to understand their psychological state. Based on the classical Chinese segmentation methods used by such dictionaries as CCIDict and CC-LIWC, this paper proposed a word segmentation algorithm that can better cover the ancient Chinese vocabulary used in imperial edicts. We used this algorithm to calculate the psycholinguistic words in imperial edicts of the Western and Eastern Jin Dynasties (265–420). We firstly collected 613 edicts from 18 emperors of the Western and Eastern Jin Dynasties, with a total word count of more than 45,000. After being analyzed and calculated by the dictionary-based classical Chinese word segmentation algorithm, all these words were divided into 78 categories of psycholinguistic words. By comparing the frequencies of such word categories in imperial edicts of the Western Jin (265–317) and the Eastern Jin (317–420), we found significant differences in the following five word categories: personal pronouns (p = 0.027), modal particles (p = 0.034), social process words (p = 0.016), difference words (p = 0.016), and time words (p = 0.043). Based on differences in these five categories, we analyzed the psychological changes of the Western Jin and Eastern Jin emperors. This paper thereby verified the applicability and feasibility of the dictionary-based classical Chinese word segmentation algorithm.
