Resumen de Big data techniques for real-time processing of massive data streams - técnicas big data para el procesamiento de flujos de datos masivos en tiempo real.
Machine learning techniques have become one of the most demanded resources by companies due to the large volume of data that surrounds us in these days. The main objective of these technologies is to solve complex problems in an automated way using data. One of the current perspectives of machine learning is the analysis of continuous flows of data or data streaming. This approach is increasingly requested by enterprises as a result of the large number of information sources producing time-indexed data at high frequency, such as sensors, Internet of Things devices, social networks, etc. However, nowadays, research is more focused on the study of historical data than on data received in streaming. One of the main reasons for this is the enormous challenge that this type of data presents for the modeling of machine learning algorithms.
This Doctoral Thesis is presented in the form of a compendium of publications with a total of 10 scientific contributions in International Conferences and journals with high impact index in the Journal Citation Reports (JCR). The research developed during the PhD Program focuses on the study and analysis of real-time or streaming data through the development of new machine learning algorithms. Machine learning algorithms for real-time data consist of a different type of modeling than the traditional one, where the model is updated online to provide accurate responses in the shortest possible time. The main objective of this Doctoral Thesis is the contribution of research value to the scientific community through three new machine learning algorithms. These algorithms are big data techniques and two of them work with online or streaming data. In this way, contributions are made to the development of one of the current trends in Artificial Intelligence.
With this purpose, algorithms are developed for descriptive and predictive tasks, i.e., unsupervised and supervised learning, respectively. Their common idea is the discovery of patterns in the data.
The first technique developed during the dissertation is a triclustering algorithm to produce three-dimensional data clusters in offline or batch mode. This big data algorithm is called bigTriGen. In a general way, an evolutionary metaheuristic is used to search for groups of data with similar patterns. The model uses genetic operators such as selection, crossover, mutation or evaluation operators at each iteration. The goal of the bigTriGen is to optimize the evaluation function to achieve triclusters of the highest possible quality. It is used as the basis for the second technique implemented during the Doctoral Thesis.
The second algorithm focuses on the creation of groups over three-dimensional data received in real-time or in streaming. It is called STriGen. Streaming modeling is carried out starting from an offline or batch model using historical data. As soon as this model is created, it starts receiving data in real-time. The model is updated in an online or streaming manner to adapt to new streaming patterns. In this way, the STriGen is able to detect concept drifts and incorporate them into the model as quickly as possible, thus producing triclusters in real-time and of good quality.
The last algorithm developed in this dissertation follows a supervised learning approach for time series forecasting in real-time. It is called StreamWNN. A model is created with historical data based on the k-nearest neighbor or KNN algorithm. Once the model is created, data starts to be received in real-time. The algorithm provides real-time predictions of future data, keeping the model always updated in an incremental way and incorporating streaming patterns identified as novelties. The StreamWNN also identifies anomalous data in real-time allowing this feature to be used as a security measure during its application.
The developed algorithms have been evaluated with real data from devices and sensors. These new techniques have demonstrated to be very useful, providing meaningful triclusters and accurate predictions in real time.
