Introduction
Socioeconomic status (SES) is a critical metric for understanding urban dynamics, enabling applications ranging from targeted urban planning to financial risk assessment. Traditionally, SES data is collected through census surveys or inferred from Call Detail Records (CDRs) and smart card data. However, these sources often lack the granularity or ubiquity required for high-precision modeling. GPS trajectories, by contrast, offer a rich, continuous stream of human mobility data. This paper presents DeepSEI (Socioeconomic-aware Deep Model), a deep learning framework designed to infer a user's SES—proxied by home property values—solely from GPS mobility records.
The DeepSEI Framework
DeepSEI addresses the inherent noise and complexity of raw GPS data by employing a two-pronged neural network architecture that captures both high-level statistical features and fine-grained sequential patterns.
1. Data Preprocessing and Feature Extraction
Raw GPS pings are first processed into "stay points," representing locations where a user remained for a significant duration. These points are then cross-referenced with Point-of-Interest (POI) data to infer the activity associated with the stay (e.g., work, recreation, or shopping). This transforms raw coordinates into a semantic sequence of human behavior.
2. Coarse-level Indicators
The first branch of the network calculates three primary indicators of mobility diversity:
- Spatiality Diversity: Measured via the "radius of gyration," this captures the physical extent of a user’s movement.
- Temporality Diversity: This uses cross-entropy to measure the regularity and predictability of a user’s daily schedule.
- Activity Diversity: This quantifies the variety of location types visited, distinguishing between specialized and varied lifestyles.
3. Detailed-level Sequential Modeling
The second branch utilizes a Hierarchical Long Short-Term Memory (LSTM) network to process the temporal sequence of activities. To manage the complexity of long-term mobility, the model is split into two layers:
- Intra-day LSTM: Captures the specific transitions and routines within a single 24-hour period.
- Inter-day LSTM: Aggregates these daily representations to identify broader weekly rhythms and periodic habits.
By concatenating the outputs of these two branches, the model generates a comprehensive profile of the user’s socioeconomic behavior.
Experimental Setup
The model was evaluated using three datasets centered on Beijing, China:
- Geolife: A longitudinal dataset containing 25 million GPS records from 182 users.
- Amap POI Data: A database of 156,000 Points-of-Interest used for semantic labeling.
- Lianjia House Price Data: Real estate transaction data from 8,124 residential locations. The house price at a user’s inferred home location served as the ground-truth proxy for their SES.
Key Insights into Mobility and Wealth
The study revealed distinct behavioral differences across socioeconomic brackets:
- The Proximity of Wealth: In the dense urban environment of Beijing, wealthier individuals tend to travel shorter distances. This suggests that high-income earners prioritize living near central business districts or work hubs to minimize commute times.
- Temporal Regularity: Higher SES users often exhibit more rigid and predictable schedules, likely reflecting stable, professional employment. Conversely, lower SES users in the study showed more fragmented and irregular mobility patterns.
Results and Performance
DeepSEI demonstrated significant improvements over existing benchmarks in both classification and clustering tasks. In binary classification (distinguishing between high and low SES), the model achieved an accuracy of 90.2% and an F1-score of 86.1%.
A notable finding from the ablation study was the importance of the sequential component; removing the Hierarchical LSTM caused the F1-score to drop to 33.4%. This confirms that the order and routine of activities are far more predictive of wealth than simple aggregate statistics like total distance traveled.
Conclusion and Future Work
DeepSEI proves that GPS mobility data is a powerful proxy for socioeconomic status. While the current model relies on house prices as a proxy—which may not account for all facets of wealth—it provides a scalable alternative to traditional surveys. Future research will explore the application of Transformer architectures for sequence encoding and investigate how these mobility patterns vary across different global urban structures, such as car-centric Western cities versus transit-oriented Asian metropolises.
Summary updated