Abstract
Beyond being facilitators of human interactions, social networks have become an interesting target of research, providing rich information for studying and modeling user’s behavior. Identification of personality-related indicators encrypted in Facebook profiles and activities are of special concern in our current research efforts. This paper explores the feasibility of modeling user personality based on a proposed set of features extracted from the Facebook data. The encouraging results of our study, exploring the suitability and performance of several classification techniques, will also be presented.
Introduction
Personality is a fundamental driver of human behavior, influencing everything from learning styles to consumer preferences. If digital systems could accurately model a user's personality, they could dynamically adapt interfaces—adjusting motivational strategies, presentation styles, or recommendations to better suit the individual.
While social media platforms like Facebook provide a wealth of behavioral data, the primary challenge lies in identifying which specific "digital breadcrumbs" correlate with personality traits. Previous research often examined social network activity (e.g., friend counts) or linguistic cues in isolation. This study bridges that gap by integrating demographic data, egocentric network metrics, and advanced natural language processing (NLP) to build robust predictive models for the "Big Five" personality traits: Openness (OPE), Conscientiousness (CON), Extraversion (EXT), Agreeableness (AGR), and Neuroticism (NEU).
Methodology
The researchers employed a multi-staged approach to feature engineering and classification using a dataset from the myPersonality project, which included 250 Facebook users and approximately 10,000 status updates.
Feature Extraction and Selection
A set of 725 features was initially generated, categorized into five groups:
- Demographics: Age and gender.
- Activity-related: Counts of likes, statuses, groups, tags, and events.
- Egocentric Network: Metrics such as network density, brokerage, and betweenness.
- Linguistics: Basic word counts (LIWC) and advanced categorization using Afinn (emotional valence) and H4Lvd (General Inquirer) for nuanced word categories like "wealth" or "enlightenment."
Pearson correlation analysis was used to identify the most relevant features for each trait. Crucially, the researchers found that using all 725 features led to poor performance. By applying aggressive feature selection and refining the models to include only the top 5–16 most relevant features per trait, classification precision improved by up to 78%.
Classification
Several algorithms were tested, with Support Vector Machines (specifically the SMO version) and Boosting algorithms (such as AdaBoost) demonstrating the highest performance.
Results
The study set a target precision of 75%. After feature selection, the SMO classifier exceeded this goal for every personality trait, achieving high True Positive (TP) rates and Area Under the ROC Curve (AUC) values.
As shown in Table 1, the model achieved its highest precision for Openness (94.8%), followed by Extraversion (92.8%) and Conscientiousness (92.0%). Even the most challenging traits to predict, Neuroticism and Agreeableness, reached precisions of 86.4% and 86.0%, respectively.
Key Predictive Indicators
- Extraversion: Linked strongly to network structure, correlating with friend counts ($r=0.4$), network density, and transitivity.
- Neuroticism: Associated with gender, network size, and the use of "social support" language.
- Conscientiousness: Correlated with event attendance and specific word categories in the H4Lvd dictionary related to authority and "reaping affect."
- Openness: Strongly predicted by geo-location data, group memberships, and language related to aesthetic skills.
Conclusion
This research demonstrates that a user's personality can be predicted with high accuracy by mining a small, highly-correlated subset of social media features. The "less is more" principle proved vital, as feature selection significantly boosted model performance. While the sample size of 250 users is a limitation, the results provide a strong foundation for future work. Next steps include expanding the dataset and moving toward practical applications, such as creating intelligent, adaptive user interfaces that respond to a user's unique psychological profile.
Summary updated