Case-Method: A Methodology for Building Large-Scale Case-Based Systems

Hiroaki Kitano, Hideo Shimazu, Akihiro ShibataAAAI 1993

Introduction

In the early 1990s, Case-Based Reasoning (CBR) emerged as a powerful paradigm for solving new problems by retrieving and adapting solutions from similar past experiences. However, while algorithmic research matured, the software engineering discipline lacked a structured methodology for deploying these systems at a corporate scale. The authors introduced CASE-METHOD, the first comprehensive software engineering methodology specifically designed for building and maintaining large-scale CBR systems. By integrating CBR with standard relational databases and iterative development cycles, they demonstrated a staggering 12x reduction in maintenance workload for corporate-wide knowledge systems.

The Challenge of Scale

Before CASE-METHOD, existing methodologies for Expert Systems—such as KADS or NEC’s EX/METHOD—were poorly suited for CBR. These traditional approaches focused on extracting rigid, static rules from human experts. In contrast, CBR is inherently data-driven and thrives on a continuous influx of new "cases." Without a structured methodology, large-scale CBR projects often devolved into unmanageable, ad-hoc efforts that failed to integrate with mainstream corporate IT infrastructure. The authors sought to bridge this gap by creating a field-tested, repeatable process for building what they termed "Experience Sharing Architectures" (ESA).

The CASE-METHOD Cycle

The core of the methodology is an iterative approach consisting of six major processes designed to manage the lifecycle of a case-based system:

  1. System Development: Utilizing standard software engineering models (like the waterfall model) to build the retrieval shell.
  2. Case-Base Development: A specialized process for collecting "seed cases," extracting attributes, and forming indexing hierarchies.
  3. System Operation: Standard deployment, user support, and data entry.
  4. Data-Base Mining: Applying statistical analysis and decision trees to extract rules and trends from the growing case base.
  5. Management: Defining organizational support and incentive systems to ensure a steady flow of high-quality cases.
  6. Knowledge Transfer: Distributing extracted knowledge across various corporate divisions.
The Case-Method Cycle illustrates the iterative processes of system development, case-base development, and management.

Technical Innovation: CARET

A key technical pillar of this methodology is CARET, an RDB-based CBR shell. Unlike traditional CBR systems that relied on proprietary memory structures, CARET maps case hierarchies directly onto standard Relational Database Management Systems (RDBMS) like Oracle. It employs a "relaxation strategy" to generate SQL queries that find "near neighbors" rather than exact matches, allowing CBR to leverage existing corporate database infrastructure without sacrificing performance.

Validation via SQUAD

The methodology was validated using SQUAD, a massive corporate-wide system at NEC for software quality control. At the time of the study, SQUAD had accumulated over 25,000 cases, with 3,000 new cases reported annually by over 15,000 active participants. The system initially launched with just 100 "seed cases" to define the starting data structure, proving that the methodology could scale from a small pilot to a massive enterprise tool.

Key Results and Insights

The implementation of CASE-METHOD yielded significant improvements in both development efficiency and system performance:

  • Workload Reduction: The most striking result was the reduction in maintenance and case-base building effort. The workload dropped from 6 man-months to just 0.5 man-months for processing 1,500 cases—a 12-fold improvement.
  • Development Speed: A system that typically required 10 man-months of development was completed in just 1.5 man-months using CASE-METHOD tools.
  • Performance: Despite using standard RDBMS, the system remained highly responsive. The average response time for a query against 1,500 cases (with 130 attributes) was approximately 2.0 seconds on a SparcStation2.
  • Data Coverage: Iterative refinement of attributes led to a 95% coverage rate, meaning nearly all new cases could be accurately described using existing system parameters.
The workload for case-base building and maintenance decreased significantly over four cycles as the number of attributes stabilized.

Beyond the metrics, the authors emphasize that CBR should be viewed as an Organizational Learning tool. By providing case reporters with a structured format derived from the system's own indexing hierarchies (the "Feedback" phase), the quality of incoming data improves while data entry costs plummet. This transforms individual "tacit" knowledge into "explicit" corporate assets.

Limitations and Future Work

While CASE-METHOD represents a major leap forward, the authors acknowledge remaining challenges. Specifically, they note the need for more robust "Verification and Validation" (V&V) methods to ensure CBR systems consistently provide correct answers. Future work also includes aligning the methodology with international standards like ISO-9000-3 and further automating the "Data-Base Mining" process through advanced machine learning techniques.

Summary updated

BibTeX
@inproceedings{kitano1993case,
  author = {Hiroaki Kitano and Hideo Shimazu and Akihiro Shibata},
  title = {{Case-Method: A Methodology for Building Large-Scale Case-Based Systems}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1993},
}
Figure