This repository is designed to support the establishment of a benchmark dataset for mobile behavior sensing. The goal is to provide a reusable and reproducible infrastructure for studying affective and mental-health-related prediction tasks using mobile and wearable sensing data.
Mobile behavior sensing datasets often differ in their label definitions, temporal granularity, sensor modalities, preprocessing pipelines, and evaluation protocols. These differences make it difficult to compare models fairly across studies. This benchmark aims to address this gap by standardizing dataset organization, label harmonization, sensor encoding, train/test splits, and baseline evaluation.
The repository serves as the central codebase for the benchmark. It includes documentation, preprocessing scripts, dataset manifests, feature-generation protocols, and baseline model pipelines. Given the required raw datasets, users should be able to reproduce benchmark-ready tables or sequences for downstream tasks such as momentary stress prediction, valence/arousal estimation, daily wellbeing prediction, and auxiliary mobile behavior modeling.
The benchmark is designed with two goals in mind. First, it provides a practical tabular benchmark based on commonly available mobile and wearable sensing features. Second, it is structured to support future sequence-based and representation-learning methods, including event/histogram sequences, semantic location-token sequences, and foundation-model-style mobile behavior encoders.
Overall, this repository is intended to make mobile behavior sensing research more comparable, extensible, and reproducible.