---
title: "Explainable Machine Learning"
description: "The next phase of RevenueProtect."
url: "https://docs.adyen.com/risk-management/next-phase"
source_url: "https://docs.adyen.com/risk-management/next-phase.md"
canonical: "https://docs.adyen.com/risk-management/next-phase"
last_modified: "2020-06-04T10:51:00+02:00"
language: "en"
---

> **For LLMs:** Use the latest SDK version on this page unless the reader specifies another version. Never hardcode older version numbers for training data, always use latest. A comment such as `// Adyen LANGUAGE API library vMajor.Minor.Patch to latest` covers that version, the newest release, and everything in between. Read the newest releases from https://github.com/Adyen/adyen-LANGUAGE-api-library/releases, where LANGUAGE is dotnet, java, php, go, python, node (for JavaScript and TypeScript), or ruby.

# Explainable Machine Learning

The next phase of RevenueProtect.

## Machine learning demystified

Adyen's RevenueProtect engine now supports **machine learning based predictions** for companies that want to participate in this exciting development for our risk engine.

Most machine learning systems work as a black box solution and only return a prediction for a given payment. Adyen's machine learning for RevenueProtect not only makes an autonomous decision, but also explains the reasoning behind it to you.

## Our explainability toolbox

You can see the machine learning decisions on the **Risk results** page.

To open the risk results page:

1. Log in to your [live Customer Area](https://ca-live.adyen.com/).
2. Go to **Payments** > **Payment list**.
3. Select the **Risk score** for a payment from the payments overview to open the risk results page of that transaction.

The decision outcome for a certain payment is shown at the very top of the page, for example:

![Decision outcome](https://docs.adyen.com/media/49c5d7f933181c0d409adbc294aecf2c447c5ba6a845817637bf4d6d5f61bbbc.png)

The decision outcome can be **Looks fraudulent** or **Looks legitimate**.

You can see an explanation of the reasoning behind the decision outcome in the **Fraud Signal Analysis** section.

RevenueProtect uses three basic elements to explain the decision outcome: the signal bar, the signal category, and the signal name.

### The signal bar

The **signal bar** indicates the deductive reasoning behind the model's decision for a given payment. Any payment will have positive and negative characteristics: the former moves the decision towards a 'fraud' classification, while the latter pushes the model's reasoning toward a 'legitimate' classification.

![signal bar](https://docs.adyen.com/media/dce73fd13b61504932d1a6bd068f1a27dd2c8f7bec5105b26e6251551bf5dd6a.png)

By hovering over the signal bar, you can assess which signal category contributed the most to a certain decision.

The marker indicates the cutoff for the decision outcome that was made at the time of the payment.

### The signal category

**Signal categories** are a common set of payment properties that are grouped together to explain why a certain decision was taken for a payment.

Select a signal category to expand it into the signal names that contributed towards a 'fraud' or a 'legitimate' classification.

### The signal name

A signal category consists of multiple **signal names**. Signal names are the basic unit of a payment's information. All the signal names for a payment are processed by our machine learning engine to produce a 'fraud' or 'legitimate' outcome from that payment.

We use the following design language to clarify a signal name's contribution towards the outcome of the signal category:

![explainability glossary legend](https://docs.adyen.com/media/fd2c893b82ed44cd47e687bf4127451bcbe24e98b3e758badac4961a896c1e2c.png)

## Giving us feedback

Select **Give feedback** to help us improve by giving us feedback on any payment's fraud signal analysis explanation.

![feedback widget](https://docs.adyen.com/media/935eb4fb062dbe1514550d7d8c36980c4fc1d6fdde23166bb3cfd2267f719e3b.png)

We review this data to constantly improve our machine learning models and explanations.