Learning to Handle Inconsistency for Multi-Source Integration

AAAI 1999

TL;DR

This paper tackles the challenge of integrating data from multiple web sources that often contain inconsistent information. They introduce a semi-automated tool that learns to map equivalent data instances across these sources, even when they are represented differently, by using a decision tree learner to identify key matching attributes. This work is crucial for building robust information integration systems that can effectively combine diverse and potentially conflicting data.

What Problem Does This Paper Address?

Integrating information from the web is a complex task, primarily because data from different sources rarely aligns perfectly. This paper focuses on a specific type of inconsistency: when the same real-world entity (like a restaurant) is described differently across various data sources. For instance, "Jerry's Famous Deli" from one source might appear as "Jerry's Famous Delicatessen" in another. Directly comparing these entries using simple equality checks fails. The challenge is to automatically identify these equivalent entries and create mappings that allow for seamless data integration, enabling systems like information brokers to provide unified and accurate information.

Methods

The authors propose a semi-automated tool with three core components:

  1. Match Generator: This component identifies potential matches between data instances from different sources. It uses general, domain-independent transformation rules (like recognizing substrings, acronyms, and abbreviations) to compare shared attributes (e.g., Name, Address, Phone). It calculates a similarity score for each potential match.
  2. Decision Tree Learner: This is the core learning component. It takes a user-labeled subset of potential matches (indicating which pairs are truly equivalent) and learns a decision tree. This tree identifies which attributes and their similarity score thresholds are most important for determining a match. For example, it might learn that the "Name" attribute's similarity score must be above 0.8 for a match to be considered.
  3. One-to-One Match Maker: This component enforces a one-to-one relationship between matched objects, which is often a realistic constraint in many application domains.

The learned decision tree is then used to classify the remaining, unlabeled potential matches.

Datasets

The paper uses a dataset of Los Angeles restaurants as an example. This dataset is derived from two sources: Yahoo's Restaurants webpage and the LA County Department of Health's website. The authors do not explicitly state the size of this dataset or if it's a newly created benchmark, but it serves as a concrete illustration of their approach.

Key Insights and Ideas

  • Beyond Equality: The fundamental insight is that simple equality checks are insufficient for data integration. Similarity-based matching, using domain-independent transformations, is essential for handling variations in data representation.
  • Learning What Matters: The decision tree learner is a powerful mechanism for automatically discovering which attributes are the most reliable indicators of a match. This avoids manual feature engineering and adapts to the specific characteristics of the data.
  • Semi-Automation is Key: While the tool is semi-automated, requiring some user-labeled data for training, it significantly reduces the manual effort compared to fully manual mapping. This balance is crucial for practical applications.

Key Results

The paper demonstrates the effectiveness of their approach using the restaurant example. They show how their tool can correctly identify matched restaurants despite variations in their names (e.g., "Jerry's Famous Deli" vs. "Jerry's Famous Delicatessen"). While specific quantitative results like accuracy percentages are not detailed in this abstract, the core contribution is the development of a functional framework that can learn to handle these inconsistencies.

Limitations and Future Work

The authors acknowledge that their current tool requires user-labeled data for training the decision tree. For future work, they propose incorporating active learning or query-by-committee techniques to reduce the size of this required training set, making the system even more efficient and scalable.

Summary updated

BibTeX
@inproceedings{1999learning,
  title = {{Learning to Handle Inconsistency for Multi-Source Integration}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1999},
}
Figure