概述
It is useful when class sizes differ and a high accuracy score could hide failure on the less common class.
深入探讨
A classifier can look successful by repeatedly choosing the common class. Imagine a test set containing 990 legitimate messages and 10 spam messages. Predicting legitimate for every message gives 99% accuracy, yet the filter catches no spam. This is a constructed example of class imbalance, not a reported performance result. MCC helps assess the relationship between predictions and outcomes using correct and incorrect classifications from both classes. In a binary confusion matrix, true positives and true negatives are correct decisions; false positives and false negatives are the two kinds of mistake. A score near positive one indicates strong agreement. Zero indicates no correlation, and negative values indicate an inverse relationship. With both binary classes present, a perfect reversal gives negative one. Consider a different hypothetical classifier with 40 true positives, 40 true negatives and 10 of each kind of error. Its accuracy is 80%, and its MCC is 0.6. Those values describe the same predictions using different scales. An MCC of 0.6 does not mean that 60% of cases were classified correctly. MCC does not decide how costly an error is. A missed machine defect and an unnecessary inspection may have very different consequences. Report the confusion matrix and relevant class-specific metrics alongside MCC, then choose an operating threshold using validation data and the actual decision costs. Scikit-learn provides matthews_corrcoef for binary and multiclass labels. The metric works on predicted classes, so it does not assess whether a claimed 90% probability is trustworthy. Degenerate cases, such as predicting only one class, also need care: the binary formula has a zero denominator, and a library's numeric convention should not be mistaken for a successful classifier.
战略影响
成本与预算
多年来，架构决策决定着性能和运营成本。
更清晰的判决
技术教育帮助团队选择正确的堆栈，而不仅仅是最新的堆栈。
质量控制
更好的工程选择可以减少生产中的可靠性事故。
The Future of Matthews Correlation Coefficient
Automated evaluation reports can make MCC more useful by linking each score to the exact dataset, label definition and decision threshold. A team should be able to move from a summary number to the cases responsible for errors, while respecting access controls on the data. Future model comparisons should also include changing class frequencies and results for relevant subgroups. No single agreement statistic can determine whether deployment is worthwhile. The practical improvement comes from combining a reproducible score with evidence about which mistakes occur and how people respond to them.
现实世界的实施
In a hypothetical set of 1,000 messages, 990 are legitimate. A filter that labels every message legitimate reaches 99% accuracy while detecting no spam, showing why accuracy needs additional context.
A toy classifier has 40 true positives, 40 true negatives, 10 false positives and 10 false negatives. Its MCC is 0.6, calculated from the full confusion matrix.
A team compares two defect detectors at fixed operating thresholds. Alongside MCC, it reports how many defective items each detector misses, because those misses have a specific operational cost.
An analyst uses scikit-learn's matthews_corrcoef on observed and predicted labels. They evaluate the model's probability confidence separately rather than treating MCC as a calibration score.
风险与防护栏
优化一项基准测试可以隐藏更广泛的系统弱点。
基础设施和维护成本常常被低估。
随着系统变得更加复杂，安全性和可观察性差距可能会扩大。
实施路线图
在实施之前定义延迟、质量和成本目标。
在实际负载和数据条件下进行基准测试。
仪器监控错误、漂移和用户影响。
在扩展之前准备回滚和事件响应路径。
不断探索
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the Matthews Correlation Coefficient quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
常见问题
What is Matthews Correlation Coefficient?
The Matthews correlation coefficient, or MCC, summarizes how well predicted class labels agree with actual labels while accounting for every cell of the confusion matrix. It is useful when class sizes differ and a high accuracy score could hide failure on the less common class.
A filter labels all 1,000 messages legitimate when 990 truly are legitimate. Why is its 99% accuracy insufficient evidence of useful spam detection?
Always selecting the majority class can produce high accuracy while missing every example of the class the filter needs to detect.
Which value results from the guide's toy confusion matrix with 40 true positives, 40 true negatives and 10 of each error?
The numerator is 1,500 and the denominator is 2,500, yielding an MCC of 0.6.
A report interprets an MCC of 0.6 as '60% of cases were correct.' How should that interpretation be corrected?
Accuracy expresses the fraction of correct predictions. MCC uses a different formula and scale.
For binary data containing both classes, a classifier reverses every label perfectly. Which MCC value describes that relationship?
A perfect inverse relationship between actual and predicted binary labels produces an MCC of negative one.
A factory values missed defects differently from unnecessary inspections. Which additional evidence is needed alongside MCC?
MCC summarizes association without encoding the specific consequences of false positives and false negatives.
继续学习
相关指南
为此主题精选的更多指南