liveBook · Manning

Bias

As explained in Machine Learning System Design

A systematic error caused by a model's limited capacity to capture useful signals, leading it toward simplified assumptions about the data.

A sales forecast can miss in two ways that call for different decisions: it may fail to learn important patterns, or it may consistently predict too high or too low. Both are called bias, but the first concerns model capacity and the second concerns the direction of forecast errors. section 9.1.1 section 12.1.2

Check which meaning matters before choosing a fix:

  • Capacity bias: a model with too little capacity misses useful relationships between features and the target. It underfits, performing poorly on both training and unseen data. Increasing capacity may reduce this bias but raises the risk of overfitting. section 9.1.1
  • Directional forecast bias: inspect signed residuals to find where predictions repeatedly miss in one direction. For product sales, promotions led to substantial underpredictions; adding a promotion calendar to the features addressed that missing signal. section 9.3.1
  • Evaluation bias: a performance estimate can be biased even when the model is not underfitting. Reusing one validation split for tuning and evaluation can produce overly optimistic, nonrobust results. Choosing more cross-validation folds can reduce estimation bias, at a cost in variance and computation. section 7.1 section 7.2.3

Equal error, different revenue

Two airline ticket-sales forecasts each have a 15% weighted average percentage error. One has −10% signed bias and tends to underpredict; the other has +10% and tends to overpredict. Under the illustrated pricing assumptions, their four-day revenues are $75,300 and $83,900. An aggregate error score alone would hide a difference that matters to the pricing decision. section 12.1.2

Match the response to the evidence: compare training with unseen performance for underfitting, examine residuals for systematic misses, and keep validation representative of how the model will be used. Bias also appears in fairness monitoring, where demographic parity and equalized odds can assess performance across groups; that is a separate question from model capacity or signed forecast error. section 9.1.1 section 9.3.1 section 7.6 section 14.4.3

How Machine Learning System Design covers it

How deep it goes
Defines capacity bias through underfitting and works through signed forecast bias with error and revenue calculations, without deriving the tradeoff mathematically.
Examples
Uses promotion-driven sales underpredictions and two airline forecasts with equal aggregate error but different signed bias and revenue.
What it assumes
Assumes familiarity with training and unseen data, features, targets, and prediction errors while defining capacity bias directly.
What you can do after
Diagnose/debug with it by inspecting residual patterns and judging whether directional forecast errors warrant different features or business decisions.
How it connects
Connects bias to variance, underfitting, overfitting, cross-validation, residual analysis, and fairness monitoring.
Cover of Machine Learning System Design
liveBookmanning.com
Read Machine Learning System Design on liveBook Get the book at manning.com

Also explained in

Compare how 28 books explain bias →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Machine Learning System Design. The section links point to the parts of the book it draws on.