The set of assumptions a learning algorithm uses to generalize from observed data to unseen instances.
Every model embodies biases - e.g., smoothness assumptions in kernel methods or locality in KNN - that guide generalization. Recognizing inductive bias helps governance teams select algorithms appropriate to the domain and understand failure modes. It also informs how much data is needed to learn reliably, and which model classes may systematically underperform on certain tasks.
A time-series team chooses an autoregressive model because its inductive bias assumes temporal continuity, fitting stock-price data better than a feedforward NN. They document this choice in their model-selection rationale for future audits and ensure the model’s bias aligns with domain knowledge.
