Abstract
Sparse mobile crowdsensing (MCS) is a promising paradigm for large-scale urban monitoring, allowing for data collection from only a few areas (cell selection) to infer the state of the entire city (data inference). While traditional Sparse MCS focuses on single variables, modern urban applications often require multi-dimensional data, such as simultaneous temperature and humidity maps. This paper presents a novel framework that jointly addresses data inference and cell selection for multi-task scenarios by exploiting intra- and inter-task correlations. By combining multi-task compressive sensing with deep reinforcement learning (DRL), the proposed system significantly reduces sensing costs while maintaining high data accuracy.
Introduction
Mobile Crowdsensing (MCS) leverages the sensors in citizens' smartphones to monitor urban environments. However, recruiting enough users to cover every subarea (cell) of a city is prohibitively expensive. Sparse MCS addresses this by collecting data from only a fraction of cells and using algorithms to fill in the gaps.
A major limitation of current research is the focus on single-variable tasks. In reality, urban planners need multi-dimensional data. Because variables like temperature and humidity or PM2.5 and PM10 are physically correlated, treating them as independent tasks is inefficient. This paper explores how to leverage these correlations to optimize both the inference of missing data and the selection of which cells to sample.
Methodology
The authors introduce a two-pronged approach to handle multi-task urban sensing: multi-task compressive sensing for inference and a reinforcement learning (RL) model for cell selection.
1. Multi-Task Compressive Sensing (Inference)
Instead of traditional matrix completion for a single task, the framework uses Collective Matrix Factorization. This allows different tasks to share a "factor matrix," enabling the model to learn cross-task relationships—for example, high temperature in a cell might imply lower humidity, even if humidity wasn't directly measured there. The model incorporates three key constraints:
- Temporal: Values change smoothly over time.
- Spatial: Nearby cells are likely to have similar readings.
- Value: Cells with similar environmental surroundings (e.g., two parks) likely have similar data profiles.
2. RL-Based Cell Selection (Optimization)
To determine which cells and tasks to prioritize, the authors utilize Deep Reinforcement Learning (DRL).
- Multi-Branch Architecture: The system uses a unique neural network where separate branches extract "intra-task" features (patterns within one variable), which are then fused to capture "inter-task" correlations (patterns across variables).
- Reward Mechanism: The RL agent is rewarded based on how much a specific selection improves the overall inference accuracy of the entire urban map.
- Two-Stage Framework: To handle the cold-start problem of RL, the authors propose an online framework. A Training Phase uses uncertainty-based sampling to gather initial data, followed by a Running Phase where the trained RL agent takes over the selection process.
Experimental Evaluation
The framework was evaluated on two real-world urban sensing datasets:
- Sensor-Scope: Environmental data (Temperature and Humidity) from 57 sensors at the EPFL campus.
- U-Air: Air quality data (PM2.5 and PM10) from 36 sensors in Beijing.
Key Results
- Efficiency Gains: The RL-assisted algorithm required 16% to 25% fewer sensed values than the state-of-the-art SPACE-TA algorithm to achieve the same target accuracy.
- Inference Accuracy: Multi-task compressive sensing consistently yielded lower error rates than single-task methods, especially in high-sparsity scenarios where very few samples were available.
- Real-time Performance: While training the RL model takes 10–30 minutes, the actual decision-making for cell selection takes less than 1ms, making it highly suitable for live deployments.
Key Insights
- Correlations as a Force Multiplier: Inter-task correlations are often stronger than spatial ones. By measuring one variable (like PM2.5), the system gains significant information about another (PM10) without needing additional sensors in that same cell.
- RL vs. Uncertainty: Most systems pick cells where the model is "most confused" (highest variance). This paper argues that being "confused" doesn't mean a cell is useful. The RL approach learns to identify "anchor cells"—roughly 30% of the grid—that are most representative of the whole system.
Conclusion
By exploiting the inherent relationships between different environmental variables, this multi-dimensional framework offers a more efficient path for large-scale urban monitoring. Future work will focus on integrating user mobility models and privacy-preserving mechanisms to make the system more robust for real-world crowdsensing deployments.
Summary updated