Technical GlossaryData Science and Data Management

Mode Collapse

A problem in synthetic data generation where the model loses distributional diversity and produces only limited types of samples.

Mode collapse is a major quality issue, especially in GAN-based generation. Instead of learning the full data distribution, the model may repeatedly generate only a limited set of modes. In that case, synthetic data may look realistic at first glance but remain weak in terms of diversity. This can mislead model training, test scenarios, and downstream applications. For that reason, synthetic data quality must be evaluated not only by sample realism, but also by distributional coverage.