March 2022 Summaries

Researchers from Gretel.ai and Illumina's Emerging Solutions have successfully created synthetic versions of real-world genomic data sets using state-of-the-art generative neural networks. The synthetic datasets offer enhanced privacy guarantees, enabling life science researchers to collaborate and test ideas through open access to data without compromising patient privacy. While the initial case study results are based on a small sample set, continued experiments in scale, accuracy, and privacy show that synthetic data has the potential to enable sharing and collaboration on synthetic genomics datasets at a much larger scale than currently possible. The code for synthesizing genomic data is available on GitHub, and further research will explore the scale and privacy guarantees achievable with synthetic data on genomic datasets.

Handling imbalanced datasets in machine learning, especially in fields such as fraud detection and cybersecurity, is challenging due to the limited instances of the minority class, like fraudulent transactions. This exploration uses a popular Kaggle dataset on credit card fraud to demonstrate how synthetic data can improve model accuracy. By employing a generative synthetic data model, the process creates additional fraudulent records by incorporating features from both fraudulent records and their nearest neighbors, labeled as non-fraudulent but potentially suspicious. This method, inspired by the Synthetic Minority Oversampling Technique (SMOTE), aims to enhance classifier performance by generating new instances that help the model generalize better to detect fraud. The approach involves using Gretel Synthetics, a tool that leverages deep learning to generate synthetic data, and optimizes the training process to balance data creation without overfitting. The addition of synthetic data to the training set aims to reduce the negative-to-positive ratio, potentially boosting the model's ability to detect fraud by up to 14%. This method underscores the potential of synthetic data to enhance machine learning models by overcoming challenges of extreme class imbalance and improving generalization across datasets.

Synthetic data is artificially annotated information generated by computer algorithms or simulations that mirrors the statistical properties of real-world datasets. It can be used as a substitute when suitable real-world data is not available, or to protect sensitive and personally identifiable information (PII) in cases where privacy concerns or compliance risks exist. Synthetic data opens up possibilities for enabling access to artificial and privacy-preserving versions of data in minutes, augmenting machine learning datasets for superior accuracy and fairness, implementing privacy-by-design principles, creating safe data retention policies, testing software products and services, training ML and AI models, sharing data within organizations, and sharing data with third parties.

Gretel AI's CPO Alex Watson discussed how synthetic data can be used for medical research while maintaining ethical, equitable and fair practices. Synthetic data is an alternative to real-world data generated by computer simulations or algorithms. It has gained popularity in recent years due to advancements in deep learning techniques. Gretel AI uses a language model that trains on sensitive customer data sets while imposing privacy parameters to prevent memorization of data it shouldn't. The resulting artificial data set maintains the same insights and distributions as the original data, but is not based on any real-world person or object. Synthetic data can be used for faster access to medical research data, reducing bias in datasets, generating more samples from limited data sets, and improving overall accuracy of machine learning models.

The discussion at The Rise of Privacy Tech’s Data Privacy Week 2022 conference addressed various questions about synthetic data, emphasizing the distinctions between synthetic data creation and differential privacy. Synthetic data is generated by algorithms that learn the dataset distribution, maintaining properties like correlations, while differential privacy introduces calibrated noise to obscure individual data points. The conversation highlighted the importance of privacy-protection mechanisms in machine learning to prevent adversarial attacks, with differential privacy being a key method. Synthetic data may not suit studies focusing on outliers or rare populations, as quality can degrade without a large dataset. Effective communication about the potential error in synthetic data is crucial, with its quality being measured based on intended use. Gretel offers synthetic data quality reports to ensure statistical fidelity, and advocates for differentially private querying systems for further understanding original datasets. The discussion concluded with an invitation for continued dialogue within the community.

This article explores the training of a FastCUT Generative Adversarial Network (GAN) model on map data and public e-bike feeds from cities across the USA. The trained GAN is then tested for its ability to learn and generalize by predicting location data sets for cities worldwide, including Tokyo. The process involves encoding e-bike location data as pixels into an image and training it as an image translation task similar to CycleGAN, Pix2pix, and StyleGAN. The newer contrastive unpaired translation (FastCUT) model is used due to its memory efficiency, fast training capabilities, and good generalization with minimal parameter tuning. The results show that the GAN can generate realistic location data for anywhere in the world, although there are some false positives, particularly in waterways.

Smart-seeding is a feature that enables synthetic data models to auto-complete partial records and text. It has been used in two challenging use cases for synthetic data: removing bias from datasets and generating synthetic time-series data. Gretel has released two new blueprints that utilize smart-seeding, automating the process of recreating time series data with seasonal patterns, reducing biases in ML datasets by equalizing minority class representation, and creating synthetic data with the same shape as original data while preserving vital portions of each row. The Time Series Data blueprint trains synthetic time-series models on trends in data, while the General Smart-Seeding blueprint allows users to take partial values from their training dataset for synthesis.