Computationally intensive, distributed and decentralised machine learning : from theory to applications

Fiosina, Jelena

Machine learning (ML) is currently one of the most important research fields, spanning computer science, statistics, pattern recognition, data mining, and predictive analytics. It plays a central role in automatic data processing and analysis in numerous research domains owing to widely distributed and geographically scattered data sources, powerful computing clouds, and high digitisation requirements. However, aspects such as the accuracy of methods, data privacy, and model explainability remain challenging and require additional research. Therefore, it is necessary to analyse centralised and distributed data processing architectures, and to create novel computationally intensive explainable and privacy-preserving ML methods, to investigate their properties, to propose distributed versions of prospective ML baseline methods, and to evaluate and apply these in various applications. This thesis addresses the theoretical and practical aspects of state-of-the-art ML methods. The contributions of this thesis are threefold. In Chapter 2, novel non-distributed, centralised, computationally intensive ML methods are proposed, their properties are investigated, and state-of-the-art ML methods are applied to real-world data from two domains, namely transportation and bioinformatics. Moreover, algorithms for ‘black-box’ model interpretability are presented. Decentralised ML methods are considered in Chapter 3. First, we investigate data processing as a preliminary step in data-driven, agent-based decision-making. Thereafter, we propose novel decentralised ML algorithms that are based on the collaboration of the local models of agents. Within this context, we consider various regression models. Finally, the explainability of multiagent decision-making is addressed. In Chapter 4, we investigate distributed centralised ML methods. We propose a distributed parallelisation algorithm for the semi-parametric and non-parametric regression types, and implement these in the computational environment and data structures of Apache SPARK. Scalability, speed-up, and goodness-of-fit experiments using real-world data demonstrate the excellent performance of the proposed methods. Moreover, the federated deep-learning approach enables us to address the data privacy challenges caused by processing of distributed private data sources to solve the travel-time prediction problem. Finally, we propose an explainability strategy to interpret the influence of the input variables on this federated deep-learning application. This thesis is based on the contribution made by 11 papers to the theoretical and practical aspects of state-of-the-art and proposed ML methods. We successfully address the stated challenges with various data processing architectures, validate the proposed approaches in diverse scenarios from the transportation and bioinformatics domains, and demonstrate their effectiveness in scalability, speed-up, and goodness-of-fit experiments with real-world data. However, substantial future research is required to address the stated challenges and to identify novel issues in ML. Thus, it is necessary to advance the theoretical part by creating novel ML methods and investigating their properties, as well as to contribute to the application part by using of the state-of-the-art ML methods and their combinations, and interpreting their results for different problem settings




Fiosina, Jelena: Computationally intensive, distributed and decentralised machine learning. from theory to applications. Clausthal-Zellerfeld 2022.


12 Monate:

Grafik öffnen


Nutzung und Vervielfältigung: