Feature Expansion for Social Media User Characterization
Year: 2021
Type: dissertation
Abstract: Personality plays an impactful role in our lives and psychologists believe that an individual’s behavior can be inferred through its personality. Recently, there have been cases of influential people in social media spreading misinformation, which is a potentially dangerous action. To prevent it, we need to identify which users will negatively impact the community, and we might be able to predict such behavior through personality recognition from their social media posts. This dissertation presents an approach to personality recognition from text. During the bibliographic revision, we learned that a text analysis tool called LIWC is repeatedly used with success for tasks of this type, thus we chose the LIWC dictionary to be the base feature set to consider. Also, we have found that Support-Vector Machine classifiers exhibit the best results. From these two findings, we outlined the following objectives: (i) exploit machine learning algorithms different from the ones used in related works to find one that produces better results; (ii) analyze whether extending LIWC’s vocabulary without supervision improves the classification results. For training and testing, we used a data set of stream-of-consciousness essays comprised of 2468 samples annotated with the Big Five personality traits of the writer: openness to experience, conscientiousness, extraversion, agreeableness, and neuroticism. We used four machine learning algorithms for classification: Support-Vector Machine, Naive Bayes, Decision Tree, and Random Forest. Also, we selected two methods for vocabulary expansion: WordNet’s synsets, and Word Embeddings. The results obtained show that the Random Forest classifier performs similarly to the algorithms used in related works, with an average accuracy of approximately 56.5%, which are promising ratings. The vocabulary expansions we have performed allowed the algorithm to match 0.6% more words from the essay data set. However, the changes to the classification results were not significant, therefore the vocabulary expansion was not beneficial.
Author: Monteiro, André Filipe da Cruz
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7047150839