On the enhancement of Big Data Pipelines through Data Preparation, Data Quality, and the distribution of Optimisation Problems

• Autores: Álvaro Valencia Parra
• Directores de la Tesis: María Teresa Gómez López (dir. tes.), Ángel Jesús Varela Vaca (dir. tes.)
• Lectura: En la Universidad de Sevilla ( España ) en 2022
• Idioma: español
• Número de páginas: 151
• Enlaces
  • Tesis en acceso abierto en: Idus
• Resumen

  • Nowadays, data are fundamental for companies, providing operational support by facilitating daily transactions. Data has also become the cornerstone of strategic decision-making processes in businesses. For this purpose, there are numerous techniques that allow to extract knowledge and value from data. For example, optimisation algorithms excel at supporting decision-making processes to improve the use of resources, time and costs in the organisation. In the current industrial context, organisations usually rely on business processes to orchestrate their daily activities while collecting large amounts of information from heterogeneous sources. Therefore, the support of Big Data technologies (which are based on distributed environments) is required given the volume, variety and speed of data. Then, in order to extract value from the data, a set of techniques or activities is applied in an orderly way and at different stages. This set of techniques or activities, which facilitate the acquisition, preparation, and analysis of data, is known in the literature as Big Data pipelines. In this thesis, the improvement of three stages of the Big Data pipelines is tackled: Data Preparation, Data Quality assessment, and Data Analysis. These improvements can be addressed from an individual perspective, by focussing on each stage, or from a more complex and global perspective, implying the coordination of these stages to create data workflows. The first stage to improve is the Data Preparation by supporting the preparation of data with complex structures (i.e., data with various levels of nested structures, such as arrays). Shortcomings have been found in the literature and current technologies for transforming complex data in a simple way. Therefore, this thesis aims to improve the Data Preparation stage through Domain-Specific Languages (DSLs). Specifically, two DSLs are proposed for different use cases. While one of them is a general-purpose Data Transformation language, the other is a DSL aimed at extracting event logs in a standard format for process mining algorithms. The second area for improvement is related to the assessment of Data Quality. Depending on the type of Data Analysis algorithm, poor-quality data can seriously skew the results. A clear example are optimisation algorithms. If the data are not sufficiently accurate and complete, the search space can be severely affected. Therefore, this thesis formulates a methodology for modelling Data Quality rules adjusted to the context of use, as well as a tool that facilitates the automation of their assessment. This allows to discard the data that do not meet the quality criteria defined by the organisation. In addition, the proposal includes a framework that helps to select actions to improve the usability of the data. The third and last proposal involves the Data Analysis stage. In this case, this thesis faces the challenge of supporting the use of optimisation problems in Big Data pipelines. There is a lack of methodological solutions that allow computing exhaustive optimisation problems in distributed environments (i.e., those optimisation problems that guarantee the finding of an optimal solution by exploring the whole search space). The resolution of this type of problem in the Big Data context is computationally complex, and can be NP-complete. This is caused by two different factors. On the one hand, the search space can increase significantly as the amount of data to be processed by the optimisation algorithms increases. This challenge is addressed through a technique to generate and group problems with distributed data. On the other hand, processing optimisation problems with complex models and large search spaces in distributed environments is not trivial. Therefore, a proposal is presented for a particular case in this type of scenario. As a result, this thesis develops methodologies that have been published in scientific journals and conferences.The methodologies have been implemented in software tools that are integrated with the Apache Spark data processing engine. The solutions have been validated through tests and use cases with real datasets.