Tensor Embedding: A Supervised Framework for Human Behavioral Data Mining and Prediction

Today’s densely instrumented world offers tremendous opportunities for continuous acquisition and analysis of multimodal sensor data, providing temporal characterization of individual behaviors. Is it possible to efficiently couple such rich sensor data with predictive modeling techniques to provide contextual and insightful assessments of individual behavior? Such data is noisy, incomplete, and collected from multiple sources, each with a different temporal resolution and dimensionality. In addition, longitudinal studies tend to examine a number of aspects of human behavior, such as well being, performance, or personality, which poses a challenge for handcrafted features engineering. To address these challenges, we propose supervised tensor embedding (STE), an algorithm for the joint decomposition of input and target variables in high-dimensional multimodal data. Latent features obtained from STE can be fed into any regression model for the estimation of the target variable(s). Following the embedding of higher-order data, features from different sources can be coupled. Additionally, we demonstrate that feature selection on higher-order data can enhance performance. The efficiency of our method is tested on two real-world datasets consisting of 29 target variables. Compared to the state of the art, we find that (i) our method outperforms others in 21 prediction tasks, and (ii) there are a few aspects of human behavior that are unpredictable regardless of the method, which can be the limitation of multilinear methods or lack of proper independent variables.