Regresar

A Tuning Approach Using Genetic Algorithms for Emergency Incidents Classification in Social Media in Ecuador

Abstract:

The social media are an excellent opportunity to extract data of almost any topic and perform different kind of analysis and design of artificial intelligence models, nevertheless, researchers need to consider different kind of relevant issues such as low quality data or the hyperparameter tuning complexity, that can affect the performance of an classification model. Twitter is a very popular social network that allows to researchers to get historical and real time data of any topic. In this paper, we proposed an approach based on a genetic algorithm to perform the hyperparameter tuning for classification models. The CRISP-DM methodology was used with six stages: problem understanding, data collection, data understanding, modeling, testing and deploy. Results show a good overall performance when a genetic algorithm was used to fit the best hyperparameters combination for a certain number of classification models, scoring more than 0.90 for all the tested models. The LSVC (Linear Support Vector Classifier) was the model that performed the best for the extracted data (+170k tweets), with a Matthews correlational coefficient (MCC) of 0.97 and 0.96 for multi-class and binary classification, respectively. From the obtained results, we can conclude that using a genetic algorithm for hyperparameter tuning of classification models was a successful alternative, and with this, a classifier model was implemented for emergency event classification for Spanish tweets from Ecuador.