Human Activity Recognition (HAR) is nowadays widely used in intelligent perception and medical detection, and the use of traditional neural networks and deep learning methods has made great progress in this field in recent years. However, most of the existing methods assume that the data has independent identical distribution (I.I.D.) and ignore the data variability of different individual volunteers. In addition, most deep learning models are characterized by many parameters and high resources consumption, making it difficult to run in real time on embedded devices. To address these problems, this paper proposes a Gate Recurrent Units (GRU) network fusing the channel attention and the temporal attention for human activity recognition method without I.I.D. By using channel attention to mitigate sensor data bias, GRU and the temporal attention are used to capture important motion moments and aggregate temporal features to reduce model parameters. Experimental results show that our model outperforms existing methods in terms of classification accuracy on datasets without I.I.D., and reduces the number of model parameters and resources consumption, which can be easily used in low-resource embedded devices.

Deep learning models have been shown to have great advantages in answer selection tasks. The existing models, which employ encoder-decoder recurrent neural network (RNN), have been demonstrated to be effective. However, the traditional RNN-based models still suffer from limitations such as 1) high-dimensional data representation in natural language processing and 2) biased attentive weights for subsequent words in traditional time series models. In this study, a new answer selection model is proposed based on the Bidirectional Long Short-Term Memory (Bi-LSTM) and attention mechanism. The proposed model is able to generate the more effective question-answer pair representation. Experiments on a question answering dataset that includes information from multiple fields show the great advantages of our proposed model. Specifically, we achieve a maximum improvement of 3.8% over the classical LSTM model in terms of mean average precision.

The node embedding method enables network structure feature learning and representation for social network community detection. However, the traditional node embedding method only focuses on a node’s individual feature representation and ignores the global topological feature representation of the network. Traditional community detection methods cannot use the static node vector from the traditional node embedding method to calculate the dynamic features of the topological structure. In this study, an incremental dynamic community detection model based on a graph neural network node embedding representation is proposed, comprising the following aspects. A node embedding model based on influence random walk improves the information enrichment of the node feature vector representation, which improves the performance of the initial static community detection, whose results are used as the original structure of dynamic community detection. By combining a cohesion coefficient and ordinary modularity, a new modularity calculation method is proposed that uses an incremental training method to obtain node vector representation to detect a dynamic community from the perspectives of coarse- and fine-grained adjustments. A performance analysis based on two dynamic network datasets shows that the proposed method performs better than benchmark algorithms based on time complexity, community detection accuracy, and other indicators.