Estudio de Detección de Depresión en Redes Sociales mediante Procesamiento de Lenguaje Natural y Aprendizaje Automático

Abstract

[EN] As many studies and indicators suggest, the prevalence of depressive disorders is on the rise in the last years. Furthermore, the use of Online Social Networks has risen on a yearly basis since their appearance and the predictions seem to suggest this trend will go on for many years. These two facts seem to be forming a thriving niche of work in which the huge data produced by users online can be used to predict certain conditions, of which mental illness seems to be one of the most promising. With the use of Natural Language Processing techniques, Machine Learning algorithms can be used as a powerful tool in this goal. This work makes an analysis of the reasons why such a tool could be useful, talking about concrete depressive disorders statistics and the use of online social networks. Then, it tries to explain how Natural Language Processing tools and Machine Learning Algorithms work and give an idea about their adequacy to the problem in question. Finally, we make a detailed comparison of the different methods designed to perform the depression detection and explain the results obtained with the different combinations and a general idea of the reasons of these results. Concretely, we were able to design some potent prediction algorithms (including Multinomial Naive Bayes, Logistic Regression, Multi Layer Perceptron and Convolutional Neural Network models) capable of correctly classifying the subjects between a depressed and control group using as input the texts posted on the famous social network Reddit, obtaining a 0.94 accuracy, 0.72 f1-score, 0.75 recall and 0.79 AUC-score.

[ES] En este trabajo se analizan y estudian distintos algoritmos de Procesamiento de Lenguaje Natural y de Aprendizaje Automático y su utilidad en la detección temprana de usuarios con depresión en Redes Sociales. Se muestran distintos tipos de procesamiento de los datos de texto junto con los diferentes algoritmos de aprendizaje supervisado utilizados para el problema de clasificación. Finalmente se comparan los resultados obtenidos con cada una de las técnicas.