False positives vs false negatives in machine learning
The key difference between a false positive and a false negative is that a false positive incorrectly asserts that something will happen, while a false negative incorrectly asserts that something will not happen. In machine learning, “the something” is usually a prediction about a category or outcome.
False positive vs negative example
Let’s image you developed a machine learning algorithm to predict whether I’m going to sip from the coffee mug in my hand within the next 15 seconds. Here’s how the false-positive vs false-negative matrix of possibilities would break down.
- It’s a true positive (TP) if you predict I will sip, and I do.
- It’s a false positive (FP) if you predict I will sip, but I don’t.
- It’s a true negative (TN) if you predict I won’t sip, and I don’t.
- It’s a false negative (FN) if you predict I won’t sip, but I do.
False positives and false negatives in real life
With any machine learning system, we always want to minimize false positives and false negatives as much as possible, but there is always a tradeoff. Even in life, we make every-day tradeoffs between false positives and false negatives
Take your email spam filter as an example.
Most email spam filters allow you to configure how aggressively spammy emails are filtered out.
- Are you willing to miss a few valid emails to have a cleaner inbox? That would be opting to favor false positives.
- Or are you willing to let more spam through to ensure valid emails don’t get mis-categorized? That would be a decision in favor of false negatives.
Deadly implications of false negatives
And there can be life-or-death implications as well. Imagine a medical screening device that detects cancer.
A false positive reported to a patient would be emotionally traumatizing. Just imagine being told you had cancer when you didn’t.
However, the consequence of a false negative in this situation is even worse. A missed diagnosis would prevent treatment in the early stages. With cancer, the result would be otherwise avoidable fatalities.
False positives and false negatives are important metrics in AI, machine learning and statistics. But they aren’t just abstract concepts. The manner in which AI architects choose to balance between the two can have real word implications.
Whether it’s filtering email, screening for disease, approving financial transactions, or flagging security threats, every predictive system balances the risk of one error against the other.
The challenge for designers, engineers, and policymakers is to understand these tradeoffs and tune systems in ways that best serve human needs, minimize harm, and preserve trust in the system.