Identificadores de fatores críticos na experiência do usuário do IDEA: uma abordagem para seleção de atributos

Abstract

Technological advances allow companies to deliver customized software to a demanding audience. And continuous monitoring and log analysis are important tools for increasing the level of user satisfaction. This work applies feature selection and machine learning techniques to identify attributes that influence the user experience in web systems. The data analyzed are logs captured by the Dynatrace tool, generated by the IDEIA system, initially composed of four tables, totaling 106 attributes and 1,544,663 records. After exploratory analysis and pre-processing, a final data set was obtained with 57 attributes and 217,255 records. to select attributes, the Filter, Wrapper and Embedded approaches were used, based on statistical and predictive criteria. In the filter, ANOVA, Chi-square and Mutual-information were applied; in Wapper, the method chosen was Reduce Feature Elimination (RFE); and, in the Embedded approach, Decision Tree, Random Forest and Gradient Boosting were used. After ranking, four most significant attributes were selected: javascriptErrorCount, visuallyCompleteTime, requestEroorCount and name. These attributes achieved a performance of 0.990819 and F1. The work demonstrated that four attributes were sufficient to obtain exceptional classification performance, even subjecting them to other collected samples. In this way, it was proven that the proposed method identified the main points that contribute to the user experience.