Using entity identification and classification for automated integration of spatial-temporal data

Ramoza Ahsan, Rodica Neamtu, Elke A. RundensteinerInternational Journal of Design & Nature and Ecodynamics 2016 · DOI

Abstract

Big data, crucial to answering economic, social, and political questions facing our society, tend to be diverse and distributed through various sites across the Internet. The creation of tools to integrate and analyze such data is of paramount interest. Yet the automation of these processes continues to be a great challenge. Our work rests on the observation that a great number of public data sources in domains ranging from economic to demographic, although of complex structure, often share key similarities, namely the presence of the Time and Location. Our proposed Data Integration through Object Modeling framework or DIOM tackles the critical problem of automating data integration from a variety of public websites by abstracting key features of multi-dimensional tables and interpreting them in the context of knowledge-centered Unified Spatial Temporal Model. Our classification-driven extractors are trained to identify and classify entities from both structured and unstructured parts of spreadsheets. The unstructured part contained in titles, headers, and footers reveals critical information, so-called Implicit Knowledge, crucial to the correct interpretation of data. Our experimental results on real world datasets from heterogeneous public data sources show increased accuracy by 25% compared to state-of-the-art approaches.

Introduction

Public data repositories, such as those maintained by the Census Bureau or the Bureau of Economic Analysis, offer invaluable insights for policy-making and economic forecasting. However, this data is often trapped in a "Wild West" of heterogeneous spreadsheet formats. The primary challenge in automating data integration lies in the structural variety of multi-dimensional tables and the presence of "implicit knowledge"—critical context located outside the structured data cells. For instance, a cell containing "500" under a "Sales" column is meaningless without the spreadsheet title indicating the units are "thousands of dollars" and the year is "2012."

Implicit and explicit knowledge in spreadsheets.

The DIOM Framework

The Data Integration through Object Modeling (DIOM) framework addresses these challenges by treating spreadsheet extraction as a classification and modeling task. DIOM maps data into a unified spatial-temporal model defined by three dimensions: Location (L), Time (T), and Metric (C).

The framework employs several key components:

  • Region & Entity Classification: Using Linear-chain Conditional Random Fields (CRF), the system labels spreadsheet segments as either the Implicit Region (unstructured text like titles and footers) or the Explicit Region (the structured table).
  • Multi-label Embedded Vectors (MLE-vectors): Each cell is assigned a vector representing its likely role (e.g., location, time, or data value).
  • Boundary Generator: Hierarchical clustering of MLE-vectors automatically detects the boundaries of Top, Left, and Value regions.
  • Data Transformation: The system reconciles the data by mapping identified entities into a structured DIOM Relation. If the implicit region specifies "amounts in thousands," the transformation module automatically scales the values to maintain data integrity.
The DIOM 3D data model and its mapping to a relational format.

Experimental Evaluation

The researchers evaluated DIOM using two primary corpora: the 2010 Statistical Abstract of the United States (SAUS), containing over 1,000 files, and a heterogeneous dataset of 250 files from sources like the National Science Foundation and the Tax Policy Center. The testing covered three domains: Economics, Taxes, and Demographics.

DIOM significantly outperformed existing strategies:

  • Accuracy: The framework showed a 25% improvement over the Automatic Spreadsheet Integrator (ASI) and a 45% improvement over Rule-Based Integrators (RBI).
  • Classifier Performance: Region and Entity classifiers achieved over 93% accuracy.
  • Boundary Detection: The system boasted an F1-score of 0.9776 for identifying table boundaries, far exceeding the 0.7450 achieved by the next best method.
  • Effort Reduction: Qualitative analysis indicated a massive reduction in manual repairs required by users.

Conclusion

The DIOM framework demonstrates that unstructured text in spreadsheets is essential metadata rather than noise. By explicitly modeling implicit knowledge and using context-aware CRF classification, DIOM automates the integration of complex spatial-temporal data with high precision. Future work will focus on reducing the human-in-the-loop requirement for training and adapting the model for other domains, such as scientific or medical datasets, which may rely on dimensions beyond Location and Time.

Summary updated

Abstract

Big data, crucial to answering economic, social, and political questions facing our society, tend to be diverse and distributed through various sites across the Internet. The creation of tools to integrate and analyze such data is of paramount interest. Yet the automation of these processes continues to be a great challenge. Our work rests on the observation that a great number of public data sources in domains ranging from economic to demographic, although of complex structure, often share key similarities, namely the presence of the Time and Location. Our proposed Data Integration through Object Modeling framework or DIOM tackles the critical problem of automating data integration from a variety of public websites by abstracting key features of multi-dimensional tables and interpreting them in the context of knowledge-centered Unified Spatial Temporal Model. Our classification-driven extractors are trained to identify and classify entities from both structured and unstructured parts of spreadsheets. The unstructured part contained in titles, headers, and footers reveals critical information, so-called Implicit Knowledge, crucial to the correct interpretation of data. Our experimental results on real world datasets from heterogeneous public data sources show increased accuracy by 25% compared to state-of-the-art approaches.
BibTeX
@article{ahsan2016using,
  author = {Ramoza Ahsan and Rodica Neamtu and Elke A. Rundensteiner},
  title = {{Using entity identification and classification for automated integration of spatial-temporal data}},
  journal = {International Journal of Design \& Nature and Ecodynamics},
  year = {2016},
  doi = {10.2495/dne-v11-n3-186-197},
}
Figure