Discovering action insights from large-scale assessment log data using machine learning

Abstract

This study introduces a novel machine learning algorithm that combines natural language processing techniques, such as Word2Vec and Doc2Vec, with neural networks to identify and validate significant actions within human action sequences. Using the 2012 Program for the International Assessment of Adult Competencies dataset, the algorithm visualizes and analyzes action sequences in a 2D vector space to uncover high-impact behaviors that influence performance. The methodology, validated across two problem sets (“Party Invitation” and “Club Membership”), successfully distinguishes performance groups by focusing on critical actions, leading to enhanced classification accuracy (up to 94.6%) and clustering coherence (silhouette score of 0.491). This approach demonstrates potential applications in personalized education, healthcare diagnostics, and consumer behavior prediction, advancing the understanding of human behavior through digital footprints.