Prediksi Kepribadian OCEAN dari Video Wajah dengan ArcFace dan LSTM

Abstract

Questionnaire-based personality assessment can be affected by self-report bias, while a single facial image cannot represent behavioral change during an interaction. Video-based facial modeling is therefore a complementary way to estimate apparent personality, rather than a substitute for psychological assessment. This study evaluates continuous prediction of the Big Five dimensions (Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism) using ArcFace transfer learning and Long Short-Term Memory (LSTM). The ChaLearn Looking at People First Impressions V2 dataset contains 10,000 clips. Ten frames were sampled uniformly from each video, faces were detected with MTCNN, and crops were resized to 112 × 112 pixels. A ResNet34-based ArcFace backbone produced a 512-dimensional embedding for each frame. Two LSTM layers summarized the temporal sequence, and a multi-output regression head predicted five scores in the range 0–1. Mean Absolute Error (MAE) was used for evaluation; the reported percentage was computed as (1?MAE)×100 and is termed regression accuracy, not classification accuracy. RMSprop achieved 90.52% on validation, while the configuration labeled baseline achieved 90.46% and RMSprop achieved 90.42% on the test split. With RMSprop, Agreeableness reached 90.85%, whereas Neuroticism reached 89.84%. These results indicate that sequential facial representations can support apparent personality estimation on this benchmark, although the differences between configurations are small. In practice, the system may support preliminary behavioral analysis in controlled settings, but facial scores alone are insufficient for high-stakes individual decisions. Replication and cross-population validation are needed before use beyond the benchmark.