Desarrollo de un modelo “bias-aware” de recomendación de contenido para usuarios de twitter con uso de estrategias de aprendizaje automático y minería de texto.

Abstract

This project proposes the development of a model that allows reducing bias in Twitter content recommendations based on bias analysis strategies. The objective is to detect, simplify, minimize, and achieve a certain degree of impartiality during the execution of algorithms that identify bias in Recommendation Systems. With the development of the bias-aware model, this work aims to evaluate and demonstrate the existence of bias. The corpus used in this project corresponds to the National Strike that occurred in June 2022 and the Crossed Death in June 2023. The procedure begins with data cleaning, employing techniques such as tokenization, removal of stopwords, elimination of duplicate tweets, among others. Once the data is cleaned, word2vec is used to create vectors of 50, 100, and 200 dimensions. Next, the k-means algorithm is applied to identify clusters. The centroids of each cluster are calculated, and once the political cluster is determined, tweets containing political content are labeled. Then, different classification algorithms such as SVM, decision trees, KNN, and Naive Bayes are evaluated for the different dimensions. Finally, a model is proposed that will use a new political dictionary formed from the minimum distances of words in the political cluster to other clusters. With this new cluster, the algorithm evaluation procedure is repeated. The metrics used include Accuracy, F1-score, Recall, Precision, F2, and F0.5