TY - GEN
T1 - Context-Aware Dual-Stream Framework for Sleep Quality Prediction Using Multi-Modal Lifelog Data
AU - Kim, Hoseong
AU - Kwon, Sujin
AU - Noh, Seungsu
AU - Cho, Sein
AU - Oh, Beomseok
N1 - Publisher Copyright:
© 2025 IEEE.
PY - 2025
Y1 - 2025
N2 - We tackle daily sleep quality prediction from multimodal smartphone and wearable lifelogs without relying on deep neural networks. We introduce a context-aware dual-stream framework that separates survey and sensor targets: surveys are modeled by a soft-voting ensemble of XGBoost, LightGBM, and CatBoost, while sensors are modeled by a single CatBoost, and the outputs are fused. The pipeline restores heart-rate gaps via IpDFT-based sinusoidal interpolation, derives GPS place semantics with DBSCAN, and integrates external weather/holiday context into day-level features. We evaluated our framework on a minute-level dataset collected from 10 participants, comprising 700 total days, using subject-independent 10-fold cross-validation. The framework achieves a macro-F1 of 0.6311, demonstrating that carefully engineered context and modality-aligned learning can outperform deep baselines while remaining compact and suitable for on-device inference.
AB - We tackle daily sleep quality prediction from multimodal smartphone and wearable lifelogs without relying on deep neural networks. We introduce a context-aware dual-stream framework that separates survey and sensor targets: surveys are modeled by a soft-voting ensemble of XGBoost, LightGBM, and CatBoost, while sensors are modeled by a single CatBoost, and the outputs are fused. The pipeline restores heart-rate gaps via IpDFT-based sinusoidal interpolation, derives GPS place semantics with DBSCAN, and integrates external weather/holiday context into day-level features. We evaluated our framework on a minute-level dataset collected from 10 participants, comprising 700 total days, using subject-independent 10-fold cross-validation. The framework achieves a macro-F1 of 0.6311, demonstrating that carefully engineered context and modality-aligned learning can outperform deep baselines while remaining compact and suitable for on-device inference.
KW - Context-Aware Computing
KW - Feature Engineering
KW - Machine Learning
KW - Multimodal Time Series
KW - On-Device Inference
KW - Sleep Quality Prediction
UR - https://www.scopus.com/pages/publications/105035072366
U2 - 10.1109/ICTC66702.2025.11388399
DO - 10.1109/ICTC66702.2025.11388399
M3 - Conference contribution
AN - SCOPUS:105035072366
T3 - International Conference on ICT Convergence
SP - 1739
EP - 1743
BT - 2025 16th International Conference on Information and Communication Technology Convergence, ICTC 2025
PB - IEEE Computer Society
T2 - 16th International Conference on Information and Communication Technology Convergence, ICTC 2025
Y2 - 14 October 2025 through 17 October 2025
ER -