An efficient and accurate method for binary quantification

• Autor(a) principal: Hassan, Waqar
• Data de Publicação: 2022
• Tipo de documento: Tese
• Idioma: eng
• Título da fonte: Biblioteca Digital de Teses e Dissertações da USP
• Texto Completo: https://www.teses.usp.br/teses/disponiveis/55/55134/tde-16082022-112629/
• topic: Aprendizado de máquina; Aprendizado supervisionado; Classificação; Classification; Distribution matching methods; Machine learning; Métodos de correspondência de distribuição; Métodos de mistura; Mixture methods; Quantificação; Quantification; Supervised learning

Quantification is a prosperous research topic that estimates the class prevalences in a test sample. Quantification and classification share several similarities. For instance, the most straightforward (yet biased) quantification method, Classify & Count (CC), utilizes a classifier to predict the class of each instance in a sample and compute the class ratios. However, CC presents a systematic error that grows as the class distribution shifts from a distribution it optimally quantifies. This issue has motivated quantification researchers to propose more reliable counting methods. Such proposals mostly outperform the CC method but are significantly more inefficient during inference. Nonetheless, the rapid inference time is vital for numerous applications. For instance, sensor data, tweets analysis, and news feeding need to process fast-paced streams or a large volume of data. This thesis investigates and proposes a highly efficient quantification algorithm capable of processing vast volumes of data typically required by Big Data and Data Stream applications. Our main technical contribution is Sample Mean Matching (SMM), a quantifier able to count a billion instances per second with state-of-the-art accuracy. Moreover, the performance of quantification methods varies with the changes in test set size, distribution shift, and balanced or imbalanced training data cases. Therefore, we integrate different setups from the literature and recommend a comprehensive experimental setup for assessing the quantifiers performances.