Injeção de ruído em serviços de aprendizagem de máquina na nuvem

Abstract

The application of Machine Learning as a Service (MLaaS) for sentiment analysis is a critical element in today’s technological landscape, impacting various areas from digital marketing to academic research. This study focused on exploring how the injection of noise into text data affects the performance of these services, using as examples the services offered by three major providers: Microsoft, Amazon, and Google. The primary goal of this research was to evaluate the robustness and reliability of these platforms against various types of noise induced through the NlpAug library. Performance metrics, particularly the F-measure, were analyzed, and the results obtained were compared among the different services. The results demonstrated a linear drop in the F-measure as the noise level in the data increased. Although Amazon and Google showed slightly superior performance compared to Microsoft, all suffered a significant performance decrease as noise levels increased. This study highlights the importance of robust pre-processing and data cleaning strategies when implementing machine learning solutions. Moreover, it was observed that certain types of noise, such as typographical errors, OCR, and random character substitution, worsens the performance of sentiment analysis models compared to others. In conclusion, this research provides essential insights into the need for a deeper understanding of the capabilities and limitations of MLaaS services in order to use them more effectively in real applications. Furthermore, it paves the way for future investigations that may explore the robustness of these services in relation to other types of data and machine learning tasks.