dataset bias

**Dataset Bias** refers to **systematic errors or skews in training data that cause models to learn unintended, misleading patterns** — the model captures the bias in the data rather than the true underlying relationship, leading to poor generalization and fairness issues. **Common Dataset Biases** - **Selection Bias**: The data is not representative of the real-world distribution — sampling is skewed. - **Label Bias**: Labels are systematically wrong for certain subgroups — annotator bias or measurement bias. - **Representation Bias**: Certain groups, conditions, or scenarios are underrepresented in the dataset. - **Measurement Bias**: The features or labels are measured differently for different subgroups. **Why It Matters** - **Fairness**: Dataset bias is the primary cause of algorithmic unfairness — biased data produces biased models. - **Generalization Failure**: Models trained on biased data fail when deployed on the true distribution. - **Semiconductor**: Training data from a single fab, tool, or time period creates bias toward those specific conditions. **Dataset Bias** is **garbage in, garbage out** — systematic data errors that cause models to learn the wrong patterns instead of the true signal.