Ensemble based on randomised neural networks for online data stream regression in presence of concept drift

The big data paradigm has posed new challenges for the Machine Learning algorithms, such as analysing continuous flows of data, in the form of data streams, and dealing with the evolving nature of the data, which cause a phenomenon often referred to in the literature as concept drift. Concept drift is caused by inconsistencies between the optimal hypotheses in two subsequent chunks of data, whereby the concept underlying a given process evolves over time, which can happen due to several factors including change in consumer preference, economic dynamics, or environmental conditions. This thesis explores the problem of data stream regression with the presence of concept drift. This problem requires computationally efficient algorithms that are able to adapt to the various types of drift that may affect the data. The development of effective algorithms for data streams with concept drift requires several steps that are discussed in this research. The first one is related to the datasets required to assess the algorithms. In general, it is not possible to determine the occurrence of concept drift on real-world datasets; therefore, synthetic datasets where the various types of concept drift can be simulated are required. The second issue is related to the choice of the algorithm. The ensemble algorithms show many advantages to deal with concept drifting data streams, which include flexibility, computational efficiency and high accuracy. For the design of an effective ensemble, this research analyses the use of randomised Neural Networks as base models, along with their optimisation. The optimisation of the randomised Neural Networks involves design and tuning hyperparameters which may substantially affect its performance. The optimisation of the base models is an important aspect to build highly accurate and computationally efficient ensembles. To cope with the concept drift, the existing methods either require setting fixed updating points, which may result in unnecessary computations or slow reaction to concept drift, or rely on drifting detection mechanism, which may be ineffective due to the difficulty to detect drift in real applications. Therefore, the research contributions of this thesis include the development of a new approach for synthetic dataset generation, development of a new hyperparameter optimisation algorithm that reduces the search effort and the need of prior assumptions compared to existing methods, the analysis of the effects of randomised Neural Networks hyperparameters, and the development of a new ensemble algorithm based on bagging meta-model that reduces the computational effort over existing methods and uses an innovative updating mechanism to cope with concept drift. The algorithms have been tested on synthetic datasets and validated on four real-world datasets from various application domains