Optimizing data augmentation to improve al model performance

• Autor(a) principal: Silva, Henrique Matheus Ferreira da
• Data de Publicação: 2021
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Biblioteca Digital de Teses e Dissertações do LNCC
• Texto Completo: https://tede.lncc.br/handle/tede/344
• topic: Data Augmentation; Small data; Algoritmos (Computação); Redes neurais (Computação); Aprendizado por computador; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO

Abstract The accuracy of Machine Learning (ML) based classification algorithms is highly dependent on the training dataset quality. However, many research topics have classification problems in which the examples distribution may vary widely, and in which specific classes may be strongly underrepresented (such as patient-specific medicine), or training data may be scarce (such as plant species classification), leading to unbalanced datasets. In both scenarios, this may result in a poor model’s efficiency. Data Augmentation Techniques try to mitigate this problem, by expanding the available training data in order to increase models performance. In this work, we present two novel techniques for data augmentation over tabular data. First, we present a method denominated SAGAD (Synthetic Data Generator for Tabular Datasets), which is based on the concept of conditional entropy. SAGAD can balance minority classes, at the same time increasing the overall size of the training set. Next, we present an extension of SAGAD for iterative learning algorithms, called DABEL (Data Generation Based on Complexity per Classes), which iteratively produces new training data samples based on class ambiguity. To validate our proposal, we simulated a small data scenario by using datasets well known in literature and also evaluated our methods on real world data. We evaluated SAGAD using four machine learning algorithms and DABEL using a neural network model. To measure our method’s performance, we developed a baseline use-case in which models are trained on small data, comparing both SAGAD and DABEL to it. We also tested other data augmentation techniques, against SAGAD. SAGAD is implemented and available via AugmenteR (S. Pereira; ferreira da silva; A.M Porto, 2021), which is an R package in CRAN, for data augmentation which currently has more than 2200 downloads.