ELM Meets Urban Big Data Analysis: Case Studies

Ningyu Zhang, Huajun Chen, Xi Chen, Jiaoyan Chen

TL;DR

This paper introduces a general framework for urban big data analysis using Extreme Learning Machines (ELM). By fusing heterogeneous data from social media and physical sensors, their approach effectively tackles challenges like data sparsity and improves prediction accuracy and efficiency in urban computing applications.

What Problem Does This Paper Address?

Urban computing involves analyzing vast amounts of diverse data from sensors, vehicles, buildings, and people to solve real-world urban issues like pollution and traffic congestion. However, existing machine learning methods struggle with the sheer volume, heterogeneity, and sparsity of this data. Specifically, combining information from different sources (e.g., text from social media vs. numerical sensor readings) and dealing with situations where data is scarce (e.g., low social media activity in smaller cities) are significant hurdles. This paper aims to bridge this gap by proposing a flexible and efficient framework that can handle these complexities.

Methods

The core of the proposed framework involves two main stages: urban knowledge fusion and ELM-based training.

  • Urban Knowledge Fusion: To handle heterogeneous data from social media (social view) and physical sensors (physical view), the authors employ a deep autoencoder [5]. This technique learns a middle-level feature representation that captures correlations between these different modalities, effectively bridging the gap between social and physical data.
  • ELM-Based Training:
    • For regions with abundant data, stacked ELMs (S-ELMs) [4] are used. S-ELMs are designed to handle large and complex datasets efficiently by serially connecting smaller ELM networks, requiring less memory.
  • For regions with sparse data, a Domain Adaptation Extreme Learning Machine (DAELM) [6] is employed. DAELM leverages knowledge from data-rich regions to compensate for data scarcity in target regions, addressing the label scarcity and data insufficiency problems.

Additionally, the paper introduces Location Based Feature Selection (LBFS) to better incorporate neighborhood information when extracting features for each urban block. This method considers factors like distance, density, and measures of surrounding areas to create more relevant feature vectors.

Datasets

The paper utilizes a variety of real-world urban datasets from China, including:

  • Social View:
    • Online business reviews from Dianping.com (covering shop ID, name, address, latitude, longitude, consumption cost, star rating, POI category, city, environment, service, overall ratings, and comments).
  • Data from Weibo and Twitter (with geotags).
  • Physical View:
    • Smart card transactions (for bus data).
  • Traffic index from nitrafficindex.com.
  • Real estate data from soufun.com.
  • Air quality data from pm25.in.
  • POI and business area data from Baidu Maps.
  • Road network data from OpenStreetMap.
  • Meteorological data from NOAA and forecast.io.

The paper does not introduce new benchmark datasets but rather applies its framework to existing urban data sources.

Key Insights and Ideas

  • Data Fusion is Crucial: Simply concatenating features from different sources is insufficient. Deep autoencoders are effective in learning shared representations that capture cross-modal correlations, leading to better performance.
  • ELM's Adaptability: The framework's ability to use different ELM variants (stacked ELM for abundant data, DAELM for sparse data) makes it highly adaptable to varying data availability across different urban regions.
  • Location Matters: Incorporating location-based feature selection (LBFS) significantly improves the relevance and predictive power of features by considering the impact of neighboring areas.
  • Urban Computing is Multi-faceted: The paper demonstrates the framework's versatility by applying it to distinct urban challenges, showcasing its potential for broad applicability.

Key Results

The paper presents results for two main case studies:

  • Smog-Related Health Hazard Prediction: The proposed ELM framework, particularly with knowledge fusion, achieved higher prediction accuracy for Public Health Index (PHI) compared to traditional methods like SVM and Random Forest. For instance, the average precision across cities was around 0.73 for the ELM framework, outperforming others.
  • Optimal Retain Store Placement: Experiments showed that the knowledge fusion approach, combined with LBFS and other techniques, significantly improved performance metrics like Normalized Discounted Cumulative Gain (NDCG@N), precision, and recall compared to simpler methods like concatenation. For Starbucks in Beijing, the "all methods" approach (including DAELM, inflection rules, LBFS, sampling, and knowledge fusion) yielded the best results. The framework also demonstrated improved efficiency, with significantly reduced training times.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that while DAELM helps with data sparsity, it still relies on transferring knowledge from abundant regions, which might not always be perfectly applicable.
  • The performance of the framework can be influenced by the quality and availability of the underlying data sources.
  • Future Work:
    • The authors plan to apply their framework to more urban computing applications.
  • They intend to investigate the distribution of their framework to handle even more massive datasets, suggesting scalability is a key area for future research.

Summary updated

BibTeX
@inproceedings{zhang2016elm,
  author = {Ningyu Zhang and Huajun Chen and Xi Chen and Jiaoyan Chen},
  title = {{ELM Meets Urban Big Data Analysis: Case Studies}},
  booktitle = {Comput. Intell. Neurosci},
  year = {2016},
  url = {https://openreview.net/forum?id=9prhN7Y6YX},
}
Figure