New Computing Paradigms of Stream Data Mining and Optimization in Non-Stationary Environments
• 大类 : 工程技术 - 2区
• 小类 : 计算机：理论方法 - 2区
Lately the number of application scenarios where fast data streams are produced with varying characteristics along time is growing at a fast pace over very diverse sectors, particularly in industrial systems (prognosis), health (condition monitoring, anomaly detection), telecommunications (ultra-fast resource allocation, fraud detection) and security (intrusion detection over high-speed communication networks) among many others. In these scenarios, data may come from devices, sensors, web sites, social media feeds, applications, and other data-intensive infrastructures and processes alike, hence they are often noisy, heterogeneous in nature and evolve over time. In this context, real-world applications require to deal with changing environments, e.g., the estimation of the best route for a fleet of transport vehicles may depend on eventual traffic jams, weather broadcast and/or the state of the highway; job shop scheduling could depend on changing requirements in the manufacturing plant; market conditions in financial models are subject to news and media.
Such circumstances pose an urgent need for developing efficient computational models for data mining (clustering, classification/regression) and optimization not only to accommodate the high rates at which data streams are delivered, but also to adapt to changes in the conditions that ultimately impact on the patterns and solutions found by such models. These cases, often referred to as online/stream analytics where data mining and optimization models should operate efficiently on dynamic (close to real-time) environments, unchain complex design challenges in their learning algorithms, as many factors need to be jointly considered such as computational complexity, accuracy/optimality, flexibility of the model to adapt to new data distributions and/or time-varying scenarios, latency requirements, etc.
This research area is a merge of topics of interest to many disparate research communities. The novelty will reside initially in how to bridge the gap between tasks of interest to these different communities, by offering hybrid dynamic approaches that are able to efficiently ingest and analyse streaming data sources produced in nonstationary environments.
This special issue focuses on such computational aspects and solicits articles dealing with online data processing models over streaming data, with an emphasis on descriptive analysis (including clustering), predictive modelling and optimization. Specifically, this special issue invites research papers to share latest research insights and present emerging results on theoretical and practical contributions related (but not limited) to:
- Dynamic optimization over time-evolving problem formulations.
- Multi-objective optimization and decision-making methods for nonstationary setups.
- Early classification over data streams.
- Semi-supervised/weakly-supervised predictive models for data stream mining.
- Unsupervised learning over data streams (e.g. clustering).
- Diversity-sensitive model construction for nonstationary concepts.
- Model adaptation to nonstationary datasets (e.g. concept drift).
- Change detection/classification approaches over evolving data streams.
- Design and validation of distributed online learning models and dynamic optimization solvers.
- Hybrid methods blending together elements from machine learning, heuristics and time series analysis.
- Computational complexity reduction strategies for learning models and optimization methods.
- New incremental models for learning/optimization.
- Model self-tuning approaches over data streams.
- Real-world applications of stream mining models and dynamic optimization solvers.