Abstract The remarkable development of human–computer interactions has created an urgent need for machines to be able to recognise human emotions. Human motions play a key role in emphasising and conveying emotions to meet the complexity of daily application scenarios, such as medical rehabilitation and social education. Therefore, this paper aims to explore hidden emotional states from human motions. Accordingly, we proposed a novel approach for emotion recognition using multiple inertial measurement unit (IMU) sensors worn on different body parts. First, the mapping relationship between emotion and human motion was established through fuzzy comprehensive evaluation, and data were collected for six emotional states: sleepy, bored, excited, tense, angry, and distressed. Second, the preprocessed data were used as input in a lightweight convolutional neural network to extract discriminative features. Third, an attention‐based sensor fusion module was developed to obtain the importance scores of each IMU sensor for generating a fused feature representation. In the recognition phase, we constructed a weighted kernel support vector machine (SVM) model with an auxiliary fuzzy function to improve the weight calculation method of kernel functions in a multiple kernel SVM. Finally, the results obtained are compared with those of similar state‐of‐the‐art studies, the proposed method showed a higher accuracy (99.02%) for the six emotional states mentioned above. These findings may promote the development of social robots with non‐verbal emotional communication capabilities.

Recently, extreme heat waves have swept the world, exacerbating the pressing issue of increased drought occurrences. Accurate water level prediction is crucial in combating droughts. However, current water level prediction models lack interchangeability between different bodies of water and ultimately fail to consider the relevance between different hydrological systems, rendering them inaccurate and ineffective. Therefore, a generalized, adaptable, and reliable water level prediction model is crucial.In this paper, we propose a meta-learning based model called MeWP to address such issues. To incorporate meta-learning methods, we initially embed and compile different lake data and subsequently generate global optimal initial parameters. We also design a scalable meta-trained learning based on dataset size to assist convergence efficiency on different water systems. The conclusions of our extensive experiments on five public datasets demonstrate the superior stability, accuracy, and flexibility of MeWP compared to other state-of-the-art methods.

Meta-learning based recommendation systems alleviate the cold-start problem through a bi-level meta-optimization process. Recommendation borrows prior experience from pre-trained static system-level parameters and fine-tunes the model in user-level for new users. However, it is more natural for the system to sample users in a dynamic online sequence in most real-world recommendation systems, which brings further challenges for existing meta-learning based recommendation: system-level updates begins before user-level recommendation models have converged on the whole time series; stable and randomness-resistant bi-level gradient descent approaches are missing in the current meta-learning framework; evaluation on learning abilities across different users are lacked for exploring the diversities of different users.