Big Five Personality Trait Classification from Social Media Text Using Naive Bayes, K-Nearest Neighbors, and Decision Tree with ADASYN-based Class Balancing

Personality classification from short user-generated text has become a useful tool for human-resource analytics, recommender systems, and computational psychology. Most public corpora that pair personality labels with social media text are heavily skewed across the five dimensions of the Big Five (OCEAN) model, which causes conventional classifiers to favour the majority class and to underestimate minority traits. In this paper we present a machine-learning pipeline that couples standard text representations with the Adaptive Synthetic (ADASYN) oversampling technique to mitigate this imbalance and to obtain stable per-trait performance. Posts collected from a publicly available Big Five dataset are normalized through tokenization, stop-word removal, and lemmatization, and are then represented as TF-IDF weighted unigram and bigram vectors enriched with simple LIWC-style category counts. ADASYN is applied to each binary trait independently before training Multinomial Naive Bayes, K-Nearest Neighbors, and Decision Tree classifiers. Experiments on five OCEAN dimensions show that ADASYN improves average accuracy by 18 to 23 percentage points over the unbalanced baseline, with Naive Bayes reaching an average accuracy of 0.802 and an average F1 of 0.79 across all five traits. The results indicate that careful class balancing can be more impactful than the choice of classifier when dealing with skewed personality corpora.