Calibration by group is a strict statistical fairness criterion ensuring that a model's confidence aligns with reality across all protected demographics. A model is calibrated by group if, for every distinct segment defined by a sensitive attribute, the proportion of positive outcomes among instances receiving a specific predicted score matches that score exactly. This prevents systematic over- or under-estimation of risk for specific populations.
Glossary
Calibration by Group
A fairness criterion ensuring that a model's predicted probabilities accurately reflect the true likelihood of an outcome for every distinct demographic group, preventing over- or under-estimation of risk.
FAIRNESS CRITERION
What is Calibration by Group?
Calibration by group is a fairness definition requiring a model's predicted probability scores to accurately reflect the true empirical likelihood of an outcome for every distinct demographic segment.
Unlike parity-based metrics that constrain decision thresholds, calibration focuses on the meaning of the score itself. A perfectly calibrated model ensures that a 10% risk prediction translates to a true 10% event rate for every group. This property is critical in high-stakes domains like credit lending or medical diagnosis, where a miscalibrated score can lead to disparate impact through distorted risk assessments.
CORE MECHANISMS
Key Characteristics
Calibration by Group ensures that a model's predicted probabilities are not just accurate on average, but are equally reliable for every distinct demographic segment.
01
Per-Group Probability Accuracy
The fundamental requirement that a model's confidence scores must reflect true empirical likelihoods within each protected subgroup. For example, if a credit model assigns a 10% default probability to 1,000 applicants in a specific demographic, approximately 100 of those applicants should actually default. This prevents systematic over-estimation (where a group is unfairly denied opportunities) or under-estimation (where a group is exposed to hidden risk).
02
Reliability Diagrams by Subgroup
A primary diagnostic tool that plots predicted probabilities against observed frequencies, segmented by group membership. A perfectly calibrated model produces a diagonal line for every group. Deviations reveal miscalibration:
- Above the diagonal: Model is under-confident, underestimating true risk.
- Below the diagonal: Model is over-confident, overestimating true risk. These diagrams make disparate miscalibration immediately visible to auditors.
03
Expected Calibration Error (ECE) Decomposition
While global ECE measures average miscalibration, group-wise ECE decomposes this error to quantify how much each demographic segment contributes. A low global ECE can mask severe offsetting errors—one group systematically over-estimated while another is under-estimated. Decomposition ensures no group's miscalibration is hidden by averaging.
04
Platt Scaling and Isotonic Regression per Group
Post-hoc recalibration techniques applied independently to each subgroup's outputs:
- Platt Scaling: Fits a logistic regression model to map raw scores to calibrated probabilities for each group.
- Isotonic Regression: Learns a non-parametric, monotonic mapping that can correct more complex distortions. Training separate calibrators per group ensures the mapping function is tailored to each segment's unique score distribution.
05
Intersectional Calibration
Extends the fairness criterion beyond single attributes to compound demographic intersections (e.g., race AND gender simultaneously). A model may appear calibrated for 'women' and for 'a specific racial group' separately, yet be severely miscalibrated for women of that specific racial group. Intersectional analysis prevents fairness gerrymandering where subgroups at the intersection of multiple protected attributes are overlooked.
06
Threshold Invariance Property
A key consequence of perfect calibration by group: the optimal decision threshold for maximizing accuracy is identical across all groups. When probabilities mean the same thing for everyone, a single threshold yields consistent expected error rates. This property decouples the fairness of the probability estimates from the downstream policy choice of where to set a decision boundary.
FAIRNESS & CALIBRATION
Frequently Asked Questions
Clear answers to the most common questions about calibration by group, its relationship to other fairness metrics, and how it's implemented in production machine learning systems.
Calibration by group is a fairness criterion requiring that a model's predicted probability scores accurately reflect the true empirical likelihood of an outcome for every distinct demographic segment. In a perfectly calibrated-by-group model, if you take all instances where the model predicts a 20% probability of a positive outcome, exactly 20% of those instances should actually experience that outcome—and this must hold true within each protected subgroup independently. The mechanism works by evaluating the calibration error per group, typically using metrics like Expected Calibration Error (ECE) or reliability diagrams stratified by sensitive attributes. Unlike demographic parity or equalized odds, calibration by group focuses on the semantic meaning of the score itself, ensuring that a risk score of 0.8 means the same thing regardless of whether the subject belongs to Group A or Group B. This is critical in high-stakes domains like credit lending, where an over-estimation of default risk for a minority group would systematically deny them access to capital.
Enabling Efficiency, Speed & Accuracy
Intelligent Analysis, Decision & Execution
We build AI systems for teams that need search across company data, workflow automation across tools, or AI features inside products and internal software.
Talk to Us
Search across company data
Give teams answers from docs, tickets, runbooks, and product data with sources and permissions.
Useful when people spend too long searching or get different answers from different systems.
Enterprise searchRAGPermissions
Read more
Automate internal workflows
Use AI to route work, draft outputs, trigger actions, and keep approvals and logs in place.
Useful when repetitive work moves across multiple tools and teams.
AI agentsWorkflow automationGovernance
Read more
Add AI to products and internal tools
Build assistants, guided actions, or decision support into the software your team or customers already use.
Useful when AI needs to be part of the product, not a separate tool.
AI integrationDecision supportModel routing
Read more
About the author
Prasad Kumkar
CEO & MD, Inference Systems
Prasad Kumkar is the CEO & MD of Inference Systems and writes about AI systems architecture, LLM infrastructure, model serving, evaluation, and production deployment. Over 5+ years, he has worked across computer vision models, L5 autonomous vehicle systems, and LLM research, with a focus on taking complex AI ideas into real-world engineering systems.
His work and writing cover AI systems, large language models, AI agents, multimodal systems, autonomous systems, inference optimization, RAG, evaluation, and production AI engineering.
LinkedIn
Limited slotsGet a Free AI Consultation
Partnered with leading AI, data, and software stack.
How We Work
Custom AI workflows for your Business
One-fit-all AI don't work for modern businesses. At Inferensys, we aim to understand your business & custom requirements; which we use to define most efficient agentic workflows, the data, and the tools for your business.
01
Review the use case
We understand the task, the users, and where AI can actually help.
Read more02
Pick the right approach
We define what needs search, automation, or product integration.
Read more03
Build the first useful version
We implement the part that proves the value first.
Read more04
Improve from there
We add the checks and visibility needed to keep it useful.
Read moreThe first call is a practical review of your use case and the right next step.
Talk to Us