Recently, more and more applications have been deployed in container, and prediction of container load is essential in Clouds for improving resource utilization and achieving service-level agreements. However, accurate prediction of container load in Clouds remains an extremely challenge because the container load fluctuates drastically at small timescales. Furthermore, with many metrics of container in Clouds, it is hard to find which metrics are going to be useful. To address these challenges, we design an efficient container load prediction model named Fisher to improve the accuracy and efficiency of prediction. It mainly includes two modules: a metrics selection module and a neural network training module. We first selects relevant metrics by metrics selection module which is a novel algorithm for shape-based time-series clustering. Afterwards, we apply a powerful deep neural network model built with bidirectional long short-term memory to predict actual load one-step-ahead. We evaluate Fisher using a 30-day load trace from a data center with 500 containers. Experiments show that Fisher can reduce training metrics while maintaining prediction accuracy. More importantly, our model significantly improves prediction accuracy over 50\% compared to other state-of-the-art methods based on auto regressive integrated moving-average and long short-term memory.

Traffic flow prediction can provide important reference data for managers to maintain traffic order, and can also be based on personal travel plans for optimal route selection. On account of the development of sensors and data collection technology, large-scale road network historical data can be effectively used, but their high non-linearity makes it meaningful to establish effective prediction models. In this regard, this paper proposes a dual-stream cross AGFormer-GPT network with prompt engineering for traffic flow prediction, which integrates traffic occupancy and speed as two prompts into traffic flow in the form of cross-attention, and uniquely mines spatial correlation and temporal correlation information through the dual-stream cross structure, effectively combining the advantages of the adaptive graph neural network and large language model to improve prediction accuracy. The experimental results on two PeMS road network data sets have verified that the model has improved by about 1.2% in traffic prediction accuracy under different road networks.

On-device learning has become a new trend for edge intelligence systems. In this paper, we investigate the on-device in-cremental learning problem, which targets to learn new classes on top of a well-trained model on the device. Incremental learning is known to suffer from catastrophic forgetting, i.e., a model learns new classes at the cost of forgetting the old classes. Inspired by model pruning techniques, we propose a new on-device incremental learning method based on weight freezing. The weight freezing in our framework plays two roles: 1) preserving the knowledge of the old classes; 2) boosting the training procedure. By means of weight freezing, we build up an efficient incremental learning framework which combines knowledge distillation to fine-tune the new model. We conduct extensive experiments on CIFAR100 and compare our method with two existing methods. The experimental results show that our method can achieve higher accuracy after incrementally learning new classes.

The issue of data sparsity poses a formidable challenge in the field of recommender systems. Encouragingly, leveraging the interactions among overlapping users in the source domain can enhance item recommendation in the target domain. The transfer of user preferences across domains is a crucial concern in the cross-domain recommendation and represents a hopeful method to address data sparsity. Most existing methods transfer users' preference information by building a preference transfer network. These methods focus on the cross-domain mapping of preference features and ignore the inherent data distribution differences between the source domain and target domain. Consequently, the mapped user embeddings do not align with the item embeddings in the target domain and the recommendation quality decreases. On this basis, we propose a new method called Adaptive Meta-Learning for Cross-Domain Recommendation (AMLCDR). The method includes a meta-learning network for fully extracting user characteristics and generating a transfer network to reduce the user preference loss, as well as a domain adaptation network to align user preference distributions. We perform comprehensive experiments to assess the efficacy of AMLCDR by utilizing a substantial real-world dataset. We validate the effectiveness of data distribution alignment in domain adaptation. For diverse cross-domain recommendation tasks under different start conditions, AMLCDR outperforms state-of-the-art models in multiple evaluation metrics.