Enterprise AI Analysis

The paper reformulates departure time prediction as a time-to-event (TTE) problem, specifically a time-to-departure (TTD) task. It uses a Transformer-based architecture to estimate survival probabilities over discretized time grids, leveraging self-attention for temporal dependencies and real-time contextual patterns. This enables efficient parallel processing and real-time updates.

Each 5-minute interval is represented by a token comprising three components: (1) contextual features from passive smartphone sensing (activity transitions, ambient environment), (2) absolute time features (interval's position in the day), and (3) day-of-week features (weekday-weekend variation). These are fused via an alpha-fusion mechanism and positional encodings.

To improve robustness, three regularization mechanisms are introduced: (1) dropout-time (randomly masks absolute time features), (2) time-scale (learnable scalar moderating temporal position embedding strength), and (3) alpha-fusion (balances contextual and temporal representations). These strategies promote generalization across diverse behavioral patterns and reduce temporal bias.

The model optimizes a discrete-time ordinal regression objective, based on negative log-likelihood for uncensored samples. It encourages high survival probability before the event and a sharp drop at the event. To mitigate sensitivity to minor timing shifts, Gaussian-Smoothed Supervision (GSS) is applied, using a normalized Gaussian weighting kernel around the departure time. Event and weekend weights further stabilize training.

The proposed model achieves the lowest Mean Absolute Error (MAE) of 2.20 hours, outperforming the best historical baseline (SVR, 2.57 hours) and context-aware classifier (iTransformer, 2.59 hours). This demonstrates the effectiveness of modeling departure as a survival process.

Kernel Density Estimates (KDE) show the proposed model's predicted departure times align closely with ground-truth distributions, unlike historical baselines that concentrate around global averages. This indicates the model's temporal adaptability, though a slight bias toward earlier predictions is observed due to the decision threshold p=0.1.

Fine-tuning the last Transformer layer and output layer with user-specific data yields modest overall MAE improvement (2.20 → 2.13 hours), with larger gains on weekends (2.07 → 1.85 hours) than weekdays (2.26 → 2.23 hours). This suggests personalized modeling improves adaptability, especially for less regular weekend patterns influenced by factors like COVID-19 schedule shifts.

DFC relies on continuous passive sensing from smartphones. User acceptance is critical, involving privacy concerns and sustained engagement. Preliminary studies using the UTAUT2 framework show privacy concerns significantly predict long-term behavioral intention. Further research focuses on privacy-aware design and ethical adoption.

To address smartphone battery drain from continuous sensing, the EVA app minimizes impact by running in the background and suspending sensing below 20% battery. For large-scale deployment, computationally intensive tasks like model inference will be offloaded to a cloud-based Battery Management System (BMS), enabling real-time decision-making while reducing on-device load.

Departure time patterns differ substantially across individuals. Users with highly variable schedules exhibit greater uncertainty, leading to higher MAE (correlation r = 0.62, p = 0.0063). Advanced personalized modeling strategies are essential to address this user-specific behavioral heterogeneity and improve prediction accuracy.