Accurate time series forecasting is crucial for optimizing resource allocation, industrial production, and urban management, particularly with the growth of cyber-physical and IoT systems. However, limited training sample availability in fields like physics and biology poses significant challenges. Existing models struggle to capture long-term dependencies and to model diverse meta-knowledge explicitly in few-shot scenarios. To address these issues, we propose MetaGP, a meta-learning-based Gaussian process latent variable model that uses a Gaussian process kernel function to capture long-term dependencies and to maintain strong correlations in time series. We also introduce Kernel Association Search (KAS) as a novel meta-learning component to explicitly model meta-knowledge, thereby enhancing both interpretability and prediction accuracy. We study MetaGP on simulated and real-world few-shot datasets, showing that it is capable of state-of-the-art prediction accuracy. We also find that MetaGP can capture long-term dependencies and can model meta-knowledge, thereby providing valuable insights into complex time series patterns.

The performance of modern database management systems highly relies on hundreds of adjustable knobs. Traditionally, these knobs are manually adjusted by database administrators, a process that is both inefficient and ineffective for tuning large-scale databases in cloud environments. Recent research has explored the use of machine learning techniques to enable the automatic tuning of database configurations. Although most existing learning-based methods achieve satisfactory results on static workloads, they often experience performance degradation and low sampling efficiency in real-world environments. According to our study, this is primarily due to a lack of safety guarantees during the configuration sampling process. To address the aforementioned issues, we propose SafeTune, an online tuning system that adapts to dynamic workloads. Our core idea is to filter out a large number of configurations with potential risks during the configuration sampling process. We employ a two-stage filtering approach: The first stage utilizes a semi-supervised outlier ensemble with feature learning to achieve high-quality feature representation. The second stage employs a ranking-based classifier to refine the filtering process. In addition, to alleviate the cold-start problem, we leverage the historical tuning experience to provide high-quality initial samples during the initialization phase. We conducted comprehensive evaluations on static and dynamic workloads. In comparison to offline baseline methods, SafeTune reduces 95.6%-98.6% unsafe configuration suggestions. In contrast with state-of-the-art methods, SafeTune has improved cumulative performance by 10.5%-46.6% and tuning speed by 15.1%-35.4%.

The widespread deployment of wireless and mobile devices results in a proliferation of spatio-temporal data that is used in applications, e.g., traffic prediction, human mobility mining, and air quality prediction, where spatio-temporal prediction is often essential to enable safety, predictability, or reliability. Many recent proposals that target deep learning for spatio-temporal prediction suffer from so-called catastrophic forgetting, where previously learned knowledge is entirely forgotten when new data arrives. Such proposals may experience deteriorating prediction performance when applied in settings where data streams into the system. To enable spatio-temporal prediction on streaming data, we propose a unified replay- based continuous learning framework. The framework includes a replay buffer of previously learned samples that are fused with training data using a spatio-temporal mixup mechanism in order to preserve historical knowledge effectively, thus avoiding catastrophic forgetting. To enable holistic representation preservation, the framework also integrates a general spatio-temporal autoencoder with a carefully designed spatio-temporal simple siamese (STSimSiam) network that aims to ensure prediction accuracy and avoid holistic feature loss by means of mutual information maximization. The framework further encompasses five spatio-temporal data augmentation methods to enhance the performance of STSimSiam. Extensive experiments on real data offer insight into the effectiveness of the proposed framework.

In the era of observability, massive amounts of time series data have been collected to monitor the running status of the target system, where anomaly detection serves to identify observations that differ significantly from the remaining ones and is of utmost importance to enable value extraction from such data. While existing reconstruction-based methods have demonstrated favorable detection capabilities in the absence of labeled data, they still encounter issues of training bias on abnormal times and distribution shifts within time series. To address these issues, we propose a simple yet effective Temporal-Frequency Masked AutoEncoder (TFMAE) to detect anomalies in time series through a contrastive criterion. Specifically, TFMAE uses two Transformer-based autoencoders that respectively incorporate a window-based temporal masking strategy and an amplitude-based frequency masking strategy to learn knowledge without abnormal bias and reconstruct anomalies by the extracted normal information. Moreover, the dual autoencoder undergoes training through a contrastive objective function, which minimizes the discrepancy of representations from temporal-frequency masked autoencoders to highlight anomalies, as it helps alleviate the negative impact of distribution shifts. Finally, to prevent over-fitting, TFMAE adopts adversarial training during the training phase. Extensive experiments conducted on seven datasets provide evidence that our model is able to surpass the state-of-the-art in terms of anomaly detection accuracy.

Cost estimation is of great importance in query optimization. However, traditional optimizers compute the cost based on heuristics, sacrificing accuracy for efficiency. In recent years, learning-based cost estimation models have achieved high accuracy. However, their poor robustness and inefficiency lead to their failure to meet the needs of practical scenarios. We propose a lightweight and Database-Agnostic Cost Estimation model (DACE) to address the above limitations. To further improve the effectiveness of DACE, we design a tree-structure-based loss adjustment strategy to learn sub-plan information and solve the information redundancy problem. As a pretrained estimator, DACE can efficiently make accurate predictions on unseen databases. For more complex scenarios, we fine-tune DACE with LoRA. The excellent efficiency allows DACE to adapt to challenging scenarios with minimal effort. As a pretrained encoder, DACE can improve the accuracy and robustness of other cost estimation models through knowledge integration and solve the notorious cold start problem. Extensive experiments have shown that DACE's accuracy, efficiency, and robustness are much better than existing methods.

Autonomous driving is an emerging technology that has developed rapidly over the last decade. There have been numerous interdisciplinary challenges imposed on the current transportation system by autonomous vehicles. In this paper, we conduct an algorithmic study on the autonomous vehicle decision-making process, which is a fundamental problem in the vehicle automation field and the root cause of most traffic congestion. We propose a perception-and-decision framework, called HEAD, which consists of an enHanced pErception module and a mAneuver Decision module. HEAD aims to enable the autonomous vehicle to perform safe, efficient, and comfortable maneuvers with minimal impact on other vehicles. In the enhanced perception module, a graph-based state prediction model with a strategy of phantom vehicle construction is proposed to predict the one-step future states for multiple surrounding vehicles in parallel, which deals with sensor limitations such as limited detection range and poor detection accuracy under occlusions. Then in the maneuver decision module, a deep reinforcement learning-based model is designed to learn a policy for the autonomous vehicle to perform maneuvers in continuous action space w.r.t. a parameterized action Markov decision process. A hybrid reward function takes into account aspects of safety, efficiency, comfort, and impact to guide the autonomous vehicle to make optimal maneuver decisions. Extensive experiments offer evidence that HEAD can advance the state of the art in terms of both macroscopic and microscopic effectiveness.

Query optimization is of great importance for the performance of a database, in which cardinality and cost estimation have a pivotal role. To enable accurate cardinality and cost estimation, we propose a novel framework based on bidirectional compressor and ensemble networks called BICE. In particular, we design a feature extractor composed of four sub-encoders, which can extract various types of information in a query plan tree and hybrid learning strategies for encoding. We encode joins based on a graph embedding method and design parallel networks for filters to improve the encoding efficiency. Then we propose a bidirectional LSTM-based compressor to learn the encoding and obtain fixed-length vectors, reducing the learning difficulty of the estimation model. Finally, we propose different data sampling strategies based on Bayesian neural networks and active learning, and an ensemble model is established based on transfer learning, which enables accurate estimation and adaptation to large-scale data queries. Extensive experiments offer insight into the effectiveness and efficiency of the proposed framework.

The proliferation of real-time monitoring applications such as Artificial Intelligence for IT Operations (AIOps) and the Internet of Things (IoT) has led to the generation of a vast amount of time-series data. To extract the underlying value of the data, both the industry and the academia are in dire need of efficient and effective methods for time-series analysis. To this end, in this paper, we propose a Multi-layer perceptron (MLP)-attention based multivariate time-series analysis model MARINA. MARINA is designed to simultaneously learn the temporal and spatial correlations among multivariate time-series. Also, the model is versatile in that it is suitable for major time-series analysis tasks such as forecasting and anomaly detection. Through extensive comparisons with the representative multivariate time-series forecasting and anomaly detection algorithms, MARINA is shown to achieve state-of-the-art (SOTA) performance in both forecasting and anomaly detection tasks.

Long sequence time-series forecasting (LSTF) has become increasingly popular for its wide range of applications. Though superior models have been proposed to enhance the prediction effectiveness and efficiency, it is reckless to neglect or underestimate one of the most natural and basic temporal properties of time series: history has inertia. In this paper, we introduce a new baseline for LSTF, named historical inertia (HI). In HI, the most recent historical data points in the input time series are adopted as the prediction results. We experimentally evaluate HI on 4 public real-world datasets and 2 LSTF tasks. The results demonstrate that up to 82% relative improvement over state-of-the-art works can be achieved. We further discuss why HI works and potential ways of benefiting from it.

Multivariate time series forecasting has been drawing increasing attention due to its prevalent applications. It has been commonly assumed that leveraging latent dependencies between pairs of variables can enhance prediction accuracy. However, most existing methods suffer from static variable relevance modeling and ignorance of correlation between temporal scales, thereby failing to fully retain the dynamic and periodic interdependencies among variables, which are vital for long- and short-term forecasting. In this paper, we propose METRO, a generic framework with multi-scale temporal graphs neural networks, which models the dynamic and cross-scale variable correlations simultaneously. By representing the multivariate time series as a series of temporal graphs, both intra- and inter-step correlations can be well preserved via message-passing and node embedding update. To enable information propagation across temporal scales, we design a novel sampling strategy to align specific steps between higher and lower scales and fuse the cross-scale information efficiently. Moreover, we provide a modular interpretation of existing GNN-based time series forecasting works as specific instances under our framework. Extensive experiments conducted on four benchmark datasets demonstrate the effectiveness and efficiency of our approach. METRO has been successfully deployed onto the time series analytics platform of Huawei Cloud, where a one-month online test demonstrated that up to 20% relative improvement over state-of-the-art models w.r.t. RSE can be achieved.