Generalizing to Unseen Domains: A Survey on Domain Generalization

IEEE Transactions on Knowledge and Data Engineering, Journal Year: 2022, Volume and Issue: unknown, P. 1 - 1

Published: Jan. 1, 2022

Machine learning systems generally assume that the training and testing distributions are same. To this end, a key requirement is to develop models can generalize unseen distributions. Domain generalization (DG), i.e., out-of-distribution generalization, has attracted increasing interests in recent years. deals with challenging setting where one or several different but related domain(s) given, goal learn model an test domain. Great progress been made area of domain for This paper presents first review advances area. First, we provide formal definition discuss fields. We then thoroughly theories carefully analyze theory behind generalization. categorize algorithms into three classes: data manipulation, representation learning, strategy, present popular detail each category. Third, introduce commonly used datasets, applications, our open-sourced codebase fair evaluation. Finally, summarize existing literature some potential research topics future.

Language: Английский

---

Spatio-Temporal Graph Neural Networks for Predictive Learning in Urban Computing: A Survey

IEEE Transactions on Knowledge and Data Engineering, Journal Year: 2023, Volume and Issue: 36(10), P. 5388 - 5408

Published: Nov. 23, 2023

With recent advances in sensing technologies, a myriad of spatio-temporal data has been generated and recorded smart cities. Forecasting the evolution patterns is an important yet demanding aspect urban computing, which can enhance intelligent management decisions various fields, including transportation, environment, climate, public safety, healthcare, others. Traditional statistical deep learning methods struggle to capture complex correlations data. To this end, Spatio-Temporal Graph Neural Networks (STGNN) have proposed, achieving great promise years. STGNNs enable extraction dependencies by integrating graph neural networks (GNNs) temporal methods. In manuscript, we provide comprehensive survey on progress STGNN technologies for predictive computing. Firstly, brief introduction construction prevalent deep-learning architectures used STGNNs. We then sort out primary application domains specific tasks based existing literature. Afterward, scrutinize design their combination with some advanced Finally, conclude limitations research suggest potential directions future work.

Language: Английский

---

Forecast evaluation for data scientists: common pitfalls and best practices

Data Mining and Knowledge Discovery, Journal Year: 2022, Volume and Issue: 37(2), P. 788 - 832

Published: Dec. 2, 2022

Abstract Recent trends in the Machine Learning (ML) and particular Deep (DL) domains have demonstrated that with availability of massive amounts time series, ML DL techniques are competitive series forecasting. Nevertheless, different forms non-stationarities associated challenge capabilities data-driven models. Furthermore, due to domain forecasting being fostered mainly by statisticians econometricians over years, concepts related forecast evaluation not mainstream knowledge among researchers. We demonstrate our work as a consequence, researchers oftentimes adopt flawed practices which results spurious conclusions suggesting methods reality be seemingly competitive. Therefore, this we provide tutorial-like compilation details evaluation. This way, intend impart information fit context ML, means bridging gap between traditional adopting current state-of-the-art techniques.We elaborate problematic characteristics such non-normality how they common pitfalls Best outlined respect steps data partitioning, error calculation, statistical testing, others. Further guidelines also provided along selecting valid suitable measures depending on specific dataset at hand.

Language: Английский

---

Deep Learning for Time Series Forecasting: Advances and Open Problems

Information, Journal Year: 2023, Volume and Issue: 14(11), P. 598 - 598

Published: Nov. 4, 2023

A time series is a sequence of time-ordered data, and it generally used to describe how phenomenon evolves over time. Time forecasting, estimating future values series, allows the implementation decision-making strategies. Deep learning, currently leading field machine applied forecasting can cope with complex high-dimensional that cannot be usually handled by other learning techniques. The aim work provide review state-of-the-art deep architectures for underline recent advances open problems, also pay attention benchmark data sets. Moreover, presents clear distinction between are suitable short-term long-term forecasting. With respect existing literature, major advantage consists in describing most such as Graph Neural Networks, Gaussian Processes, Generative Adversarial Diffusion Models, Transformers.

Language: Английский

---

State of charge estimation for lithium-ion batteries based on cross-domain transfer learning with feedback mechanism

Journal of Energy Storage, Journal Year: 2023, Volume and Issue: 70, P. 108037 - 108037

Published: June 21, 2023

Language: Английский

---

Foundation Models for Time Series Analysis: A Tutorial and Survey

Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Journal Year: 2024, Volume and Issue: unknown, P. 6555 - 6565

Published: Aug. 24, 2024

Time series analysis stands as a focal point within the data mining community, serving cornerstone for extracting valuable insights crucial to myriad of real-world applications. Recent advances in Foundation Models (FMs) have fundamentally reshaped paradigm model design time analysis, boosting various downstream tasks practice. These innovative approaches often leverage pre-trained or fine-tuned FMs harness generalized knowledge tailored analysis. This survey aims furnish comprehensive and up-to-date overview While prior surveys predominantly focused on either application pipeline aspects they lacked an in-depth understanding underlying mechanisms that elucidate why how benefit To address this gap, our adopts methodology-centric classification, delineating pivotal elements time-series FMs, including architectures, pre-training techniques, adaptation methods, modalities. Overall, serves consolidate latest advancements pertinent accentuating their theoretical underpinnings, recent strides development, avenues future exploration.

Language: Английский

---

Spatio-Temporal Graph Few-Shot Learning with Cross-City Knowledge Transfer

Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Journal Year: 2022, Volume and Issue: unknown, P. 1162 - 1172

Published: Aug. 12, 2022

Spatio-temporal graph learning is a key method for urban computing tasks, such as traffic flow, taxi demand and air quality forecasting. Due to the high cost of data collection, some developing cities have few available data, which makes it infeasible train well-performed model. To address this challenge, cross-city knowledge transfer has shown its promise, where model learned from data-sufficient leveraged benefit process data-scarce cities. However, spatio-temporal graphs among different show irregular structures varied features, limits feasibility existing Few-Shot Learning (\emph{FSL}) methods. Therefore, we propose model-agnostic few-shot framework called ST-GFSL. Specifically, enhance feature extraction by transfering knowledge, ST-GFSL proposes generate non-shared parameters based on node-level meta knowledge. The nodes in target city via parameter matching, retrieving similar characteristics. Furthermore, reconstruct structure during meta-learning. reconstruction loss defined guide structure-aware learning, avoiding deviation datasets. We conduct comprehensive experiments four speed prediction benchmarks results demonstrate effectiveness compared with state-of-the-art

Language: Английский

---

Forecasting movements of stock time series based on hidden state guided deep learning approach

Information Processing & Management, Journal Year: 2023, Volume and Issue: 60(3), P. 103328 - 103328

Published: March 1, 2023

Language: Английский

---

Neural network-based parametric system identification: a review

International Journal of Systems Science, Journal Year: 2023, Volume and Issue: 54(13), P. 2676 - 2688

Published: Aug. 2, 2023

Parametric system identification, which is the process of uncovering inherent dynamics a based on model built with observed inputs and outputs data, has been intensively studied in past few decades. Recent years have seen surge use neural networks (NNs) owing to their high approximation capability, less reliance prior knowledge, growth computational power. However, there lack review network modelling paradigm parametric particularly time domain. This article discussed connection principle between conventional models three types NNs including Feedforward Neural Networks, Recurrent Networks Encoder-Decoder. Then it reviewed advantages limitations related research addressing two major challenges interpretability nonstationary realisations. Finally, new future trends network-based identification are presented this article.

Language: Английский

---

Deep Time Series Forecasting Models: A Comprehensive Survey

Mathematics, Journal Year: 2024, Volume and Issue: 12(10), P. 1504 - 1504

Published: May 11, 2024

Deep learning, a crucial technique for achieving artificial intelligence (AI), has been successfully applied in many fields. The gradual application of the latest architectures deep learning field time series forecasting (TSF), such as Transformers, shown excellent performance and results compared to traditional statistical methods. These applications are widely present academia our daily lives, covering areas including electricity consumption power systems, meteorological rainfall, traffic flow, quantitative trading, risk control finance, sales operations price predictions commercial companies, pandemic prediction medical field. learning-based TSF tasks stand out one most valuable AI scenarios research, playing an important role explaining complex real-world phenomena. However, models still face challenges: they need deal with challenge large-scale data information age, achieve longer ranges, reduce excessively high computational complexity, etc. Therefore, novel methods more effective solutions essential. In this paper, we review developments TSF. We begin by introducing recent development trends then propose new taxonomy from perspective neural network models, comprehensively articles published over past five years. also organize commonly used experimental evaluation metrics datasets. Finally, point current issues existing suggest promising future directions combined This paper is comprehensive related years will provide detailed index researchers those who just starting out.

Language: Английский

---