A Systematic Analysis of Fine-Grained Human Mobility Prediction with On-Device Contextual Data

User mobility prediction is widely considered to be helpful for various sorts of location based services on mobile devices. A large amount of studies have explored different algorithms to predict where a user will visit in the future based on their current and historical contexts and trajectories. Most of them focus on specific targets of predictions, such as the next venue a user checks in or the destination of her next trip, which usually depend on what their task is and what is available in their data. While successful stories are often reported, little discussion can be found on what happens if the prediction targets vary: whether coarser locations are easier to be predicted than finer locations, and whether predicting the immediate next location on the trajectory is easier than predicting the destination. On the other hand, commonly used in these prediction tasks, few have utilized finer grained, on-device user behavioral data, which are supposed to be indicative of user intentions. In this paper, we conduct a systematic study on the problem of mobility prediction using a fine-grained real-world dataset. Based on a Markov model, a recurrent neural network, and a multi-modal learning method, we perform a series of experiments to investigate the predictability of different types of granularities of prediction targets and the effectiveness of different types of signals. The results provide many insights on what can be predicted and how, which sheds light on real-world mobility prediction in general.

💡 Research Summary

The paper “A Systematic Analysis of Fine‑Grained Human Mobility Prediction with On‑Device Contextual Data” conducts a comprehensive empirical study on mobile user movement prediction using a large‑scale, fine‑grained dataset collected from smartphones. Unlike most prior work that focuses on a single prediction target (e.g., the next point‑of‑interest) and a fixed set of features (historical locations, timestamps, POI tags), this study investigates three orthogonal dimensions that can fundamentally affect prediction performance: (1) location granularity, (2) target salience (operationalized as staying time), and (3) the inclusion of on‑device behavioral signals (app usage, system status, sensor events, etc.).

Problem formulation and research questions
The authors formalize mobility prediction as predicting a “target location” (l_t) from a user’s historical trajectory (T_{u}^{w_h}) and a feature set (F). The target is defined as the first future location whose salience exceeds a criterion (C). They pose three research questions (RQs): RQ1 – how does location granularity (G) affect predictability? RQ2 – how does target salience (stay duration) influence accuracy? RQ3 – what is the predictive power of multiple behavioral features, and does it vary across targets?

Dataset
The dataset comprises three months of logs from 1,200 Android users, amounting to over 200 million records. Each record includes high‑resolution GPS/Wi‑Fi location (≈10 m accuracy), timestamps, and a rich set of device‑side signals (12 categories, e.g., app launch, screen state, battery level, network type). Trajectories are extracted by splitting continuous location streams whenever a gap larger than five minutes occurs.

Methodology
The authors design a four‑step analysis pipeline: (1) data preprocessing, (2) granularity analysis, (3) salience analysis, and (4) behavioral feature analysis. For each step they train three families of models: a first‑order Markov chain, a Long Short‑Term Memory (LSTM) recurrent neural network, and a multimodal fusion model that combines heterogeneous features via separate embeddings followed by an attention‑based fusion layer. All experiments use 5‑fold cross‑validation and bootstrap resampling to assess statistical significance.

Findings – Granularity
Three granularity levels are defined: coarse (city‑district, average area ≈ 10 km²), medium (neighborhood, ≈ 1 km²), and fine (building/indoor zone, ≈ 0.01 km²). The Markov model achieves high accuracy on coarse granularity (Top‑1 ≈ 71 %), but performance drops sharply for medium (≈ 44 %) and fine (≈ 22 %) levels due to sparse transition matrices. The LSTM consistently outperforms Markov on medium and fine granularity (Top‑1 ≈ 58 % and 48 % respectively), demonstrating its ability to capture longer‑range dependencies. The multimodal model yields modest additional gains (≈ 3‑5 % absolute) over LSTM by leveraging auxiliary signals.

Findings – Salience
Salience is quantified by staying time (S). Three thresholds are examined: 5 min, 15 min, and 30 min. As the threshold rises, the set of “meaningful” target locations shrinks but becomes more predictable. For fine granularity, predicting locations with (S ≥ 30 ) min raises LSTM Top‑1 accuracy from 48 % to 68 %, indicating that users exhibit more regular patterns when they linger. The Markov model shows a similar trend but with lower absolute performance.

Findings – Behavioral Features
Baseline models using only location records achieve the scores reported above. Adding behavioral features improves performance, especially for fine granularity and high‑salience targets. The most impactful feature groups are (i) app‑switch frequency, (ii) screen‑on/off patterns, and (iii) battery consumption trends, which together contribute an average 6‑9 % absolute increase in Top‑1 accuracy. The multimodal fusion model, which learns attention weights over feature embeddings, consistently outperforms single‑modality variants, achieving the highest reported Top‑5 accuracy (≈ 85 % for coarse, 71 % for fine).

Design implications

1. Choose granularity according to service needs – coarse granularity can rely on simple probabilistic models; fine granularity requires sequence models and richer features.
2. Target salience matters – defining the prediction target as a location with sufficient dwell time yields substantially higher predictability; applications should align their objectives (e.g., destination recommendation vs. waypoint alerts) with appropriate salience thresholds.
3. Leverage on‑device context – even lightweight behavioral signals collected locally can significantly boost prediction, especially when privacy‑preserving on‑device inference is desired.
4. Multimodal fusion is beneficial – integrating heterogeneous signals via attention‑based fusion provides the most robust performance across all settings.

Conclusion
The paper delivers the first systematic, large‑scale evaluation of how problem setup (granularity, salience, and feature set) influences human mobility prediction. By demonstrating that finer granularity and higher salience increase difficulty, yet can be mitigated with sequential deep models and on‑device contextual data, the work offers concrete guidance for researchers and practitioners building location‑based services. The findings encourage future work to consider adaptive model selection and dynamic feature engineering based on the specific prediction horizon and application constraints.