Dataset Overview
Purpose & Scope
EmoPhone is a longitudinal dataset designed to advance research in emotion sensing, affective computing, and digital mental health. Data were collected from university students across three annual collection waves (D-1 in 2020, D-2 in 2021, and D-3 in 2022), each running approximately four weeks and capturing continuous behavioral signals in naturalistic, everyday settings.
Passive smartphone sensing streams — including mobility patterns, app usage, screen behavior, and communication logs — are combined with wearable physiological data to form a rich multimodal signal base. All sensing is passive and continuous, avoiding the burden of active interaction beyond scheduled EMA prompts.
Self-reported affective labels collected through ecological momentary assessment (EMA/ESM) provide ground truth for supervised learning tasks. The dataset supports benchmarking across several generalization axes, including temporal, cross-user, and cross-wave evaluation scenarios.
Data Collection
Participants were recruited from the same university-student population under an IRB-approved protocol, provided informed consent, installed the ABC Logger app on their own Android phones, and wore a designated Fitbit HR device throughout the study period. Each wave ran approximately four weeks, with realized observation spans of 22.5 (D-1), 31.7 (D-2), and 30.2 (D-3) days per person on average. ESM prompts were sent 15, 12, and 10 times per day across the three waves (an intentional reduction), yielding 53,139 total ESM responses. Quality control retained 92, 99, and 106 participants per wave based on ESM response rate and smartphone/wearable data coverage thresholds.
Label Types
A shared dimensional label core — valence, arousal, stress, and disturbance — is collected across all three waves and forms the benchmarkable label set. The third wave (D-3) adds an expanded PANAS-style affect inventory with eight richer descriptors: happy, relaxed, cheerful, content, sad, angry, anxious, and depressed, enabling finer-grained emotion analysis. For binary classification, continuous ESM responses are converted into user-relative HIGH/LOW targets using a user-specific mean split after scale harmonization.