Labels vs tags

Even with the richest real-time signals, risk engines can only go so far in accuracy detection. Without a robust feedback mechanism, even the best-in-class models will plateau, never learning the nuances of your business, your customers or your evolving fraud tactics. That’s where labels come in.

Unlike simple tags, labels feed verified “ground truth” back into SEON’s AI and machine learning engine, enabling our models to continuously refine their understanding of what’s truly legitimate and what’s fraudulent.

Models trained on live, business-specific outcomes stay razor-sharp. Labels capture both the decision on an alert or transaction (approve or decline) and the reason behind it, giving the AI and machine learning engine the detailed context it needs to uncover complex patterns and emerging fraud tactics.

By harnessing a wide spectrum of real?time signals and precise, label-driven feedback, models continually improve fraud detection accuracy, minimize false positives and directly protect revenue while enhancing customer experience.

Labels are verified outcomes that note the final “good” or “bad” verdict on each alert and transaction based on downstream confirmations. Each alert and transaction can carry a single label that precisely captures the root reason.

Below are examples of generic labels that can be applied through initial label automation in lieu of transaction states.

• Positive label: Marked as approved
• Negative label: Fraud 

If it is possible to unearth more of the root cause, more specific labels can be applied.

• Positive labels: Verified payment, On-time payment
• Negative labels: Bonus abuse, Chargeback fraud, Multi-accounting, Second payment default

When a transaction is processed through the scoring engine in SEON, it automatically triggers all applicable rules. These rules shape the final risk score assigned to the transaction and determine the transaction’s initial state: Approve, Review or Decline.

But this is only a preliminary assessment. After completing your manual review, you can apply labels to record your final decision, which will be used to train our algorithms.

1. Connect to the Label API

Implement SEON’s Label API v2 to push verified labels in bulk or real-time. Labels overwrite initial transaction states (e.g. approve, review or decline), directly influencing both the AI insights score and rule suggestions.

2. Select the applicable labels you plan to use

In order to manually adjust labels when you are reviewing alerts or transactions, you need to first select all the labels you plan to use that are applicable to your use cases. To do this, you need to navigate to AI & Machine Learning Settings, select Label API v2 and turn on access to the labels you wish to use.

3. Pick the most applicable label for each transaction

After you have reviewed a transaction or alert, adjust the transaction state to apply a more specific label use case than the generic positive and negative labels that are applied automatically. This will give the model more clarity and detail about the specific outcome of your transaction, so it can continuously learn and improve automatic labeling.

4. Downstream impacts of labels

Labels feed into AI and machine learning models as explicit “positive” or “negative” signals based on the use case. Your applied rules, AI Insights score and rule suggestions all adapt based on this feedback loop.

Learn more: Learn how to connect to the Label API and read the technical documentation.

The AI Insights score uses labels to refine its probability scoring (0–100) and can fully automate decisions above or below set thresholds. The model also translates patterns into human-readable AI-based rule recommendations, complete with accuracy metrics you can deploy directly in SEON.

1. Automate labeling to enable continuous feedback loops

Automate label ingestion via the Label API V2 immediately after your systems confirm an outcome. Faster label flow accelerates model retraining on new fraud patterns.

2. Use precise labels

Use the most specific label available (e.g., bonus abuse, account takeover) rather than a generic “fraud” tag. Granularity yields fine-tuned, high-precision rules and predictions.

3. Balanced feedback

Add labels to both approval and decline scenarios because models require a mix of good and bad examples to learn effectively. If you leave your model starved of legitimate transaction labels, it may skew overly conservative.

Approved by rules, flagged in review

A transaction passes your rule-based automation and is approved; however, due to alert trigger or automated actions you have set up, it could be flagged for review. While initially it may be labeled as Marked as approved, however, after investigation, it could be determined to apply a more specific label such as Account takeover. After this new label is applied, the model will learn this feedback and apply this label to future transactions as applicable, despite the rules that were triggered.  

Declined by rules, verified legitimate

Rules fire on an unusual velocity pattern, but analysts confirm it’s genuine. Label Marked as approved so models reduce false positives over time.

Review state ambiguity

For “Review” states, apply the label that most precisely captures the final decision, ensuring the model knows whether to reinforce “approve” or “decline” in similar edge cases.

Labels are not to be confused with tags, as they serve different purposes. Labels inform models to continuously learn and refine outcomes and improve risk detection over time.

Tags do not inform algorithms and are used to classify or add context to transactions to enable advanced filtering capabilities or for use in rules.

Use labels to train your models — use tags to organize and act on your data.

1. Connect to the Tag API or use the “Tags” column in the transactions or alert details page.

2. Add custom tags (up to 100 total, 50 at once).

3. Use tags in velocity rules to allow for more granular and precise rule configurations.

4. Automate tagging after you review an alert or transaction so that any transaction or alert containing the same use case will automatically apply the same tags.

Learn more: Learn how to connect to the Tag API and read the technical documentation.

By embedding labels into your feedback loops, you empower SEON’s models to become smarter, faster and more aligned with your unique fraud profile, delivering the agility and precision today’s fraud leaders and technical product owners demand.