Resumen de Learning the impact of data pre-processing in data analysis

Besim Bilalli

• There is a clear correlation between data availability and data analytics, and hence with the increase of data availability --- unavoidable according to Moore's law, the need for data analytics increases too. This certainly engages many more people, not necessarily experts, to perform analytics tasks. However, the different, challenging, and time consuming steps of the data analytics process, overwhelm non-experts and they require support (e.g., through automation or recommendations).

  A very important and time consuming step that marks itself out of the rest, is the data pre-processing step. Data pre-processing is challenging but at the same time has a heavy impact on the overall analysis. In this regard, previous works have focused on providing user assistance in data pre-processing but without being concerned on its impact on the analysis. Hence, the goal has generally been to enable analysis through data pre-processing and not to improve it.

  In contrast, this thesis aims at developing methods that provide assistance in data pre-processing with the only goal of improving (e.g., increasing the predictive accuracy of a classifier) the result of the overall analysis.

  To this end, we propose a method and define an architecture that leverages ideas from meta-learning to learn the relationship between transformations (i.e., pre-processing operators) and mining algorithms (i.e., classification algorithms). This eventually enables ranking and recommending transformations according to their potential impact on the analysis.

  To reach this goal, we first study the currently available methods and systems that provide user assistance, either for the individual steps of data analytics or for the whole process altogether. Next, we classify the metadata these different systems use and then specifically focus on the metadata used in meta-learning. We apply a method to study the predictive power of these metadata and we extract and select the metadata that are most relevant.

  Finally, we focus on the user assistance in the pre-processing step. We devise an architecture and build a tool, PRESISTANT, that given a classification algorithm is able to recommend pre-processing operators that once applied, positively impact the final results (e.g., increase the predictive accuracy). Our results show that providing assistance in data pre-processing with the goal of improving the result of the analysis is feasible and also very useful for non-experts. Furthermore, this thesis is a step towards demystifying the non-trivial task of pre-processing that is an exclusive asset in the hands of experts.