56% of Classification Types Lack Training Data

Across four available dark pattern datasets containing 5,561 instances, only 30 of 68 taxonomy types were represented, a 44% coverage rate.

38 dark pattern types (56%) have zero instances in any available dataset. Without training examples, ML-based detection tools cannot learn to identify these patterns.

The data imbalance is severe: “Low Stock” and “Small Close Button” patterns each have 600+ instances, while six types have fewer than 10 instances each.

Related: 05-atom—detection-coverage-gap