Banca de DEFESA: RODRIGO RIBEIRO CAPUTO

Deep learning applied to structured data like OLAP cubes

-

The Database Knowledge Discovery (KDD) process applied to structured data is a challenge where several approaches have achieved relative success for some time. Many of these approaches are based on traditional Machine Learning and its paradigms. The processing of unstructured data, on the other hand, had its popularity only recently, with the emergence of Deep Learning, an approach based on Artificial Neural Networks (ANNs). Although ANNs are not a method new, the evolution of computational power in an accessible way, especially through GPUs, together with the emergence of increasingly efficient optimization techniques, boosted its applications in an extraordinary way in domains such as image processing, audio, natural language, among others. This work seeks to investigate how these new techniques can be applied in order to improve the results in models generated on structured data. From this investigation, OlapNet's proposal arose, a Convolutional Neural Network (CNN) architecture that is based on implicit data cubes as input. Formally it was identified that this architecture is capable of surpassing the results of a specific group of transformations carried out on the data, allowing in part the automation of the data transformation stage in the KDD process. In order to verify the proposal empirically, a sample of data from a real database was used containing anonymized data on the historical indebtedness of clients of a financial institution. Based on this data, a predictive classification problem was modeled to estimate the likelihood that any customer will contract new credits in the next three months. Thus, traditional learning methods were used Machine and CNN variations, including the proposal of this work. The results showed that in almost all cases the CNNs outperform traditional methods, indicating that the Feature Maps generated from the convolutional kernels learned by the network are capable of extracting relevant characteristics. These kernels not only allow you to extract characteristics, they also reduce the complexity of the network by delimiting a neighborhood for each pixel and decrease the propensity to occur overfitting. Among the CNNs tested, OlapNet outperforms all other methods, indicating that the proposed architecture is very promising.