Filtragem colaborativa em pesquisas de clima organizacional: predição de índice de favorabilidade e de ocorrência de comentários

Abstract

Collaborative Filtering (CF) can be summarized as the process of predicting users preferences and deriving useful patterns by studying their activities. In this work, CF is used to predict the level of favorability and the occurrence of comments in answers to questions of large organizational climate surveys of a company. We aim to compare the performance of four algorithms based on CF (item-item, matrix factorization, logistic matrix factorization and neural collaborative filtering) and a baseline approach represented by a simple average of scores. The algorithms are used to estimate responses of low favorability, i.e., those that a respondent does not agree with a positive statement about the company. In addition, the algorithms are also used to estimate the registration of optional comments of respondents. For both problems, data from different checkpoints are used, comprising altogether more than 1.25 million employees’ responses. The data was collected from 2019 to 2021 by a large Brazilian company of technology with more than 10,000 employees. The results show that collaborative filtering approaches provide relevant alternatives for discriminating low favorability answers in the Likert scale as well as the occurrence of comments, with good quality estimates in both cases. These results can be further explored to eventually reduce the size of the questionnaires, avoiding burden phenomena faced by respondents when dealing with large surveys.