Recurrent neural network models have problems such as memory loss and gradient disappearance when dealing with long time series data. This paper proposes a long-term trajectory prediction model based on Transformer to process long-term sequence information. Firstly, the position encoding is used to preserve the relative positional relationship between trajectory points. Secondly, the multi-head attention mechanism is used to fully learn the feature information between different trajectories, and the trajectory data can be encoded at one time. Finally, the encoder and decoder mechanism is used to predict future trajectory data. Compared with the long-term trajectory prediction benchmark method TrajAirNet, the average displacement error, absolute displacement error of the proposed model on the long-term trajectory dataset are reduced by about 8.2% and 51.4%, respectively. The experimental results show that the proposed model has higher accuracy and robustness on long-term trajectory prediction dataset.

The performance of software fault localization techniques is critical to software debugging and the reliability of software. Spectrum-based fault localization (SBFL) and mutation-based fault localization (MBFL) are the two most popular fault localization methods. However, the accuracies of the two methods are still limited. For example, only 10.63% of faults can be detected by inspecting the top 3 suspicious elements reported by Ochiai, which is a famous SBFL technique. Unfortunately, programmers only examine the first few suspicious elements before losing patience. Since the information used in SBFL and MBFL are quite different and complementary, this paper proposes a novel approach by combining spectrum and mutation to improve the fault localization accuracy. First, the faulty program is evaluated by using SBFL, and the potential faulty statements are ranked according to their suspiciousness. Then, mutants of the program are generated and executed by MBFL. Finally, the statements that are ranked in the top tied n by SBFL are evaluated and reranked according to their mutation scores. Experiments are carried on the Defects4J benchmark and the results reveal that the accuracy of the proposed approach outperforms those of the SBFL and MBFL techniques. In terms of the faults located by inspecting the top 1 suspicious elements, the SMFL techniques detect at least 2.36 times more faults than two SBFL techniques (DStar and Ochiai) and detect at least 1.86 times more faults than two MBFL techniques (MUSE and Metallaxis).

Few-shot learning (FSL) is a core topic in the domain of machine learning (ML), in which the focus is on the use of small datasets to train the model. In recent years, there have been many important data-driven ML applications for intrusion detection. Despite these great achievements, however, gathering a large amount of reliable data remains expensive and time-consuming, or even impossible. In this regard, FSL has been shown to have advantages in terms of processing small, abnormal data samples in the huge application space of intrusion detection. FSL can improve ML for scarce data at three levels: the data, the model, and the algorithm levels. Previous knowledge plays an important role in all three approaches. Many promising methods such as data enrichment, the graph neural network model, and multitask learning have also been developed. In this paper, we present a comprehensive review of the latest research progress in the area of FSL. We first introduce the theoretical background to ML and FSL and then describe the general features, advantages, and main methods of FSL. FSL methods such as embedded learning, multitask learning, and generative models are applied to intrusion detection to improve the detection accuracy effectively. Then, the application of FSL to intrusion detection is reviewed in detail, including enriching the dataset by extracting intermediate features, using graph embedding and meta-learning methods to improve the model. Finally, the difficulties of this approach and its prospects for development in the field of intrusion detection are identified based on the previous discussion.

Information used in existing ontology matching solutions are usually grouped into four categories: lexical information, structural information, semantic information, and external information, respectively. By summarizing and analyzing the approaches for utilizing the same kind of information, this paper finds that lexical information is mainly analyzed based on text and dictionary similarity. Similarly, structural information and semantic information are mainly analyzed based on graph structure and reasoner, respectively. The approaches for aggregating information analysis results are discussed. Challenges in the analysis of various types of information for existing ontology matching solutions are also described, and insights into directions for future research are provided.

It is important to minimize the energy consumption of large-scale, geographically distributed edge data centers (EDCs). While modern processing units (PUs) have energy-saving features like Dynamic Voltage and Frequency Scaling (DVFS) and Per-Core Power Gating (PCPG), optimization is still complex and requires a holistic approach. This paper presents a new decentralized, three-timescale, online optimization approach that enables multicore micro data centers (MDCs) to optimize their per-PU power states, per-enabled-PU voltage-frequency levels and offloading schedules at three different timescales. The key idea is that we employ multi-timescale Lyapunov optimization to decouple the energy minimization between workload scheduling and result delivery at a small timescale and PU configuration at large timescales. Another important aspect is that we apply the primal decomposition to decouple the PU configuration between a per-enabled-PU voltage-frequency level at an intermediate timescale and a per-PU power state at a large timescale. Experiments demonstrate that the proposed approach improves energy efficiency significantly by up to 4.5 times in our considered lightly loaded situations where DVFS alone does not work effectively, compared to existing benchmarks.

The mobile social network contains a large amount of information in a form of commentary. Effective analysis of the sentiment in the comments would help improve the recommendations in the mobile network. With the development of well‐performing pretrained language models, the performance of sentiment classification task based on deep learning has seen new breakthroughs in the past decade. However, deep learning models suffer from poor interpretability, making it difficult to integrate sentiment knowledge into the model. This paper proposes a sentiment classification model based on the cascade of the BERT model and the adaptive sentiment dictionary. First, the pretrained BERT model is used to fine‐tune with the training corpus, and the probability of sentiment classification in different categories is obtained through the softmax layer. Next, to allow a more effective comparison between the probabilities for the two classes, a nonlinearity is introduced in a form of positive‐negative probability ratio, using the rule method based on sentiment dictionary to deal with the probability ratio below the threshold. This method of cascading the pretrained model and the semantic rules of the sentiment dictionary allows to utilize the advantages of both models. Different sized Chnsenticorp data sets are used to train the proposed model. Experimental results show that the Dict‐BERT model is better than the BERT‐only model, especially when the training set is relatively small. The improvement is obvious with the accuracy increase of 0.8%.