Best practices

Choosing Synthesis Scenarios

Most existing tabular data synthesis algorithms focus on algorithmic development, with few offering complete business solutions. When facing the complexity and incompleteness of real-world data, customization by consulting teams is often required for different application domains.

In light of this, since 2024, the CAPE team has been assisting Taiwan’s public enterprises and financial institutions in implementing synthetic data applications, developing a methodology from practical experience. We share our practical insights and demonstrate how to utilize PETsARD to address the most common and critical real data patterns, aiming to provide valuable references for data science and privacy protection teams both domestically and internationally.

Best Practices

Multi-table and Multi-timestamp Data Synthesis: A Corporate Data Case Study

• Collaborated with a policy-based financial institution to synthesize corporate customer data (including basic information, financing applications, and financial tracking), enabling them to host datathon competitions that invite external vendors to tackle business challenges.
• The dataset spans multiple business tables with complex relationships, containing multiple key timestamps from company establishment through financing applications to financial tracking, exhibiting clear temporal characteristics and business logic relationships
• The case study demonstrates how to effectively handle multi-table and multi-timestamp data through denormalization and time anchoring methods, ensuring synthetic data preserves the business logic and temporal relationships of the original data
• This best practice is applicable to processing data in similar scenarios such as corporate financing, loan applications and tracking, especially when dealing with business data that has complex table structures and involves multiple time nodes

Categorical Variables and High-Cardinality Variables Synthesis: A Higher Education Case Study

• Collaborated with a public university to synthesize student enrollment and academic performance data (including schools/departments, admission channels, course selections, etc.) to support educational policy and socioeconomic research, promoting fair academic resource access while addressing privacy concerns
• The dataset contains highly sensitive personal information such as students’ socioeconomic background, ethnicity, and disabilities, with these privacy attributes exhibiting complex and intricate potential logical relationships with students’ admission status, academic choices, and learning performance
• The case demonstrates how two methods—Uniform Encoding and Constraint Conditions—effectively process high-dimensional discrete attributes and categorical combinations, ensuring synthetic data preserves the distribution of sensitive characteristics and complex logical relationships of the original data
• This best practice is applicable to similar scenarios such as census data, social attitude surveys, consumer behavior and product portfolio research, behavioral and traffic trajectory data, especially when maintaining complex dependencies between high-dimensional categorical variables is essential

Synthesizing Low-Cardinality Data: Social Services Data Case Study (WIP)

• Collaborating with a municipal social protection service agency to synthesize cross-institutional (social affairs, police, medical) assessment and intervention questionnaires, covering initial evaluations and follow-up visits
• The dataset primarily consists of yes/no questions, single-choice, and multiple-choice questions, characterized by few options and uneven response distributions
• This best practice applies to similar low-cardinality data scenarios, such as market research surveys, user experience studies, public opinion polls, and socioeconomic statistics, particularly when dealing with structured questionnaires with standardized options

Synthesizing Imbalanced Data: Insurance Data Case Study (WIP)

• Collaborating with a Taiwanese financial holding group to synthesize insurance policy, claims, and medical visit data from its life insurance subsidiary, supporting cross-enterprise fraud detection model development
• The dataset’s key target variable is claims review results, with rejected claims accounting for only 3%, representing a typical class imbalance case
• This best practice applies to similar imbalanced data scenarios, such as credit card fraud detection, cybersecurity threat identification, and anomalous transaction screening, particularly when handling highly skewed target distributions