TL;DR
This paper introduces DeepSEI, a novel deep learning model that infers users' socioeconomic status (SES) from their GPS mobility records. By analyzing spatial, temporal, and activity patterns at both coarse and detailed levels, DeepSEI achieves superior performance compared to existing methods, opening doors for applications like risk assessment in financial services.
What Problem Does This Paper Address?
While mobility data has been extensively used for inferring demographic attributes, predicting future movements, and understanding urban dynamics, a significant gap exists in inferring users' socioeconomic status (SES). Existing methods often rely on manual data collection or are limited in scope. However, SES is crucial for various real-world applications, such as assessing creditworthiness for loans. This paper aims to bridge this gap by leveraging readily available GPS mobility data to infer SES, using house prices as a proxy. The challenge lies in extracting meaningful SES indicators from the complex and often noisy patterns of human movement.
Methods
The core of this paper is the DeepSEI model, a supervised deep learning framework designed to infer SES from mobility records. DeepSEI comprises two main components:
- Data Preprocessing: This stage involves cleaning raw GPS trajectories by filtering noise, detecting "stay points" (indicating activities like working or resting), and inferring the semantic meaning of these stay points by associating them with Point-of-Interest (POI) categories.
- Deep Network: This component captures coarse-grained features from mobility data, focusing on three aspects:
- Spatiality Diversity: Measured by the radius of gyration, indicating the spatial range of a user's activities.
- Temporality Diversity: Quantified using cross-entropy on stay durations, reflecting the regularity of a user's daily routines.
- Activity Diversity: Calculated using cross-entropy on movement patterns between locations, capturing the variety of activities a user engages in.
- Recurrent Network: This component captures detailed, sequential patterns in user activities. It utilizes a hierarchical LSTM (Long Short-Term Memory) network. The low-level LSTM models daily activity sequences, while the high-level LSTM models inter-day patterns. This hierarchical structure helps manage long sequences and preserve periodic information. The recurrent network takes spatial, temporal, and semantic features of activities as input.
The outputs from both the deep network and the recurrent network are concatenated and fed into a fully connected layer with a softmax function for final SES classification. The model is trained end-to-end, with separate pre-training for the deep and recurrent networks to provide a warm start.
Datasets
The experiments were conducted using three primary datasets:
- Mobility Data: The Geolife dataset was used, containing GPS trajectories from users over five years, primarily generated in Beijing, China. This dataset is publicly available and anonymized.
- POI Data: Point-of-Interest data for Beijing was collected via the Amap Map API. This data includes venue names, categories (e.g., working, residence, food and drink), and location information. The POI data was collected for the period corresponding to the Geolife mobility records.
- House Price Data: This dataset was collected by crawling online housing agents (Lianjia) for residential sale prices in Beijing. Each record includes sale price, floor size, and residential address. This data serves as a proxy for users' socioeconomic status.
Key Insights and Ideas
- Mobility as a SES Indicator: The fundamental idea is that where and how people move reveals significant information about their socioeconomic standing. This is a departure from solely relying on demographic attributes.
- Hierarchical LSTM for Temporal Patterns: Using a hierarchical LSTM effectively captures both daily routines and longer-term temporal regularities in user movements, which are crucial for inferring SES.
- Multi-faceted Feature Extraction: The DeepSEI model's strength lies in its ability to extract features from mobility data from multiple perspectives: spatial extent, temporal regularity, and activity diversity, at both coarse and fine-grained levels.
- House Price as a Proxy: Using house prices as a proxy for SES is a practical approach, especially given data availability and privacy concerns. The paper validates this by showing a strong correlation between house prices and income.
- Richer Users Travel Shorter: A surprising finding from the case study suggests that richer users tend to have smaller spatiality diversities, implying they travel shorter distances. This is hypothesized to be due to being busier with work.
- Secure Jobs and Regularity: The study infers that richer users often have secure jobs, leading to more regular and predictable mobility patterns (lower temporal/activity diversity).
Key Results
- Superior Performance: DeepSEI consistently outperformed baseline methods (SES, DIF, L2P) in both classification and clustering tasks. For binary classification, it achieved a 22.5% improvement in F1-score over the best baseline, and for clustering, a 37.9% improvement in ARI.
- Ablation Study Highlights Importance: Removing either the deep network or the recurrent network significantly degraded performance, emphasizing the necessity of both components. Spatiality diversity was found to be the most impactful feature among the deep network's inputs.
- Parameter Sensitivity: The model's performance is sensitive to parameters like stay point duration (optimal at 90 minutes) and cell size for spatial features (optimal at 200m).
- Case Study Validation: Visualizations and feature analysis of four users (two richer, two poorer) provided qualitative evidence supporting the model's insights, such as richer users having smaller spatial diversities and higher regularity.
Limitations and Future Work
- Proxy Limitations: The use of house prices as a proxy for SES, while practical, might not capture all nuances of socioeconomic status.
- Data Granularity: The study acknowledges that the effectiveness of features can be affected by granularity settings (e.g., spatiality granularity).
- Future Work: The authors suggest exploring more mining and learning tasks based on mobility records, such as anomaly detection. They also mention that other sequence encoder models, like Transformers, could be explored in future work.
Summary updated