Desarrollo de dos modelos de clasificación usando boosting y redes neuronales para codificar las actividades económicas y ocupaciones de investigaciones sociodemográficas del INEC.

Abstract

In statistical offices and institutions that collect data on the labor characteristics of individuals, the textual variables economic activity and occupation are coded. This activity serves to facilitate data processing and generate relevant indicators for government planning. At the National Institute of Statistics and Census of Ecuador (INEC), coding is carried out by people trained for this activity. This type of coding, called manual coding, requires a large number of people and can take a long time depending on the amount of data. For example, in the 2010 Population and Housing Census, coding was performed by 310 people lasting 5 months. Because of this, in this master's thesis two classification models were developed to automatically code the economic activities and occupations of sociodemographic research of INEC. For the development of the models, we used Xgboost and Artificial Neural Networks of Feedforward and LSTM type algorithms. The best performing models were obtained using Feedforward Neural Networks, with an accuracy of 95.18% for economic activity and 86.85% for occupations. Compared to manual coding, the implementation of these models to automatically code around 15,000 economic activities and occupations, allowed the reduction time from days to minutes. Furthermore, considering a combined approach (automatic and manual), in which the error rate of the models was less than 5%, the time was reduced to a quarter regarding manual coding and the number of personnel to half.