Troubleshooting High amounts of False Positives when detecting DATE_TIME entities.

[SpectrumSftwr]

Hi all,

I’m looking for some guidance, particularly around handling with the DATE_TIME entity.

We’re currently seeing a high volume of false positives that appear to be coming from the SpacyRecognizer. In many cases, numeric values in our input data are being incorrectly classified as DATE_TIME, which is creating noise and impacting downstream processing.

For context, our current pipeline sends masked numeric values as part of our PII detection workflow. This has led us to question whether this approach is contributing to the issue.

example: {[#]}/{[#]}/{[#]} for a date like 4/22/26

I’m curious what the general consensus or best practice is here:

• Is it better to send pseudo-numbers (synthetic but realistic numeric patterns) to preserve structure?
• Or should we continue sending fully masked values, even if it reduces context for the recognizers?

Confidence Threshold Behavior

Another area we’re struggling with is confidence scoring. We’re observing that both false positives and true positives are consistently returning confidence scores around ~0.85, which makes it difficult to use thresholding as an effective filter.

Has anyone encountered similar behavior where confidence scores don’t provide meaningful separation between correct and incorrect detections? If so:

Did you adjust or recalibrate thresholds successfully using configurations to the model?
Are there recommended approaches for tuning confidence scoring in Presidio?

Additionally, has anyone addressed similar false positives specifically tied to SpacyRecognizer? If so, did you mitigate this via:

• Custom recognizers or regex tuning
• Adjusting confidence thresholds
• Disabling or modifying SpaCy-based detection
• Pre-processing input differently

On another note about we have a threshold set to 0.85 confidence,

Any insights, trade-offs, or examples would be really helpful. Happy to share sample inputs if that would add clarity.

Thanks in advance!

[omri374]

Hi,

You’re running into a couple of known limitations with the default setup, especially around the SpaCy-based recognizer.

1) DATE_TIME false positives & masking strategy

First, regarding the false positives: the SpaCy recognizer is designed to be fast, not highly precise. It’s a general-purpose NER model, and DATE_TIME is one of the noisier entity types, especially when the input loses context (If I understand your comment on masked numeric values).

A few practical directions you can take:

1. Use one of the other NERs that are built into Presidio like a transformers model or GliNER.

2. Disable DATE_TIME in SpaCy and replace it with a custom pattern for dates you are expecting to have in your data (there is a simple date recognizer in Presidio you could start from)

2) Confidence scores (~0.85 everywhere)

What you’re seeing here is expected behavior. SpaCy does not output real confidence scores for entities. Therefore, in Presidio, the ~0.85 score is essentially a fixed/default value, not a calibrated probability. In this case, thresholding won’t help much as there's no meaningful separation between true/false positives. This is a known limitation and has come up multiple times and only applies to spaCy (other NER models output confidence scores you could threshold)

I'm not sure I fully understood your setting with the masked numeric values, so apologies if these suggestions don't fully align to your needs.

[SpectrumSftwr]

Thank you for your suggestions this was exactly what I was looking for.