Outcomes overview

From Genesys Documentation
Jump to: navigation, search

Genesys Predictive Engagement's AI-powered outcome scoring service learns to predict your business outcomes using machine learning models that are unique to your business.

Predicting outcome scores

Genesys Predictive Engagement uses AI to predict whether a visitor will achieve a particular business objective. These business objectives are called outcomes. For each outcome you define, there is a machine learning model that evaluates a visitor's behavior against the behavior of other visitors on your website to determine the visitor's outcome score. Your organization's models are unique to you.

Each outcome score represents the likelihood that the visitor will achieve the particular business outcome, based on the actions the visitor has taken so far during the session or on other activities related to the visitor that are included in the appropriate events (for example, geolocation).

The model updates a visitor's score for each outcome in real-time, and a visitor's scores can change as they navigate your website.

If a visitor communicates with an agent, the agent can see the visitor's outcome scores while viewing the visitor's complete set of journey context data. Also, the outcome scores can trigger action maps that enhance a visitor's engagement on your site.
Because our internal service handles the training and deployment of models, it is not possible to deploy customer-specific models with Genesys Predictive Engagement. Our team investigates and integrates new algorithms into the scoring service based on customer use cases.

AI-148 Outcome score updates

Previously, the Outcome Score Service used a lightweight statistical approach to predict the probability of outcome achievement based on the pages a visitor visits. Each night, the Outcome Score Service analyzed visits to your website from the previous 24 hours to learn the patterns of page views that are most predictive of each outcome. While very effective, the Outcome Score Service only considered page view events.

Now, the Outcome Score Service uses “sequence learning.” For each outcome configured, the Outcome Score Service manages a Machine Learning (ML) model that is trained exclusively on your organization’s customer journey data and tailored to your business objective. Each “sequence learning” model considers the full sequence of web events in the customer journey. These events include page views, form fills, searches, and your organization’s custom events. An event contains diverse attributes related to a customer (such as their geolocation and device type) and their website activity (such as search keywords). You can enrich events further with custom attributes that capture the unique properties of your data.

The Outcome Score Service transforms sequences of web events into an ML-friendly format and determines which features have the most predictive power. It uses customer journey data sampled from the previous week to train an ML model that identifies the patterns that are more predictive of outcome achievement. The Outcome Score Service monitors the performance of each model continuously, and retrains models regularly on the latest data. This process allows the system to react quickly to emerging trends and adapt to changes in the data.

By considering the full customer journey and managing a unique ML model tailored to each outcome, the Outcome Score Service can now identify more insightful patterns of behavior and provide a higher level of predictive power. Action maps backed by sequence learning are an even more powerful tool for offering the right engagement at the right moment.

How Genesys Predictive Engagement gathers outcome probability data

Genesys Predictive Engagement monitors all the ways that visitors arrive on and interact with your webpages. For example, if you are an e-commerce site, Genesys Predictive Engagement tracks how visitors navigate your website and place items in their shopping cart as they proceed to the checkout page.

  • How a visitor achieves a certain probability score is unique to your business and website.
  • Outcome scores and the associated data science are scored in a GDPR-compliant manner. Predictive Engagement's data scientists work exclusively on anonymized GDPR-compliant data.
For more information, see About the data we track.

Start training your models

Each of your machine learning models require training before they can make predictions.

To start the training:

  1. Create an outcome.
  2. Have visitors use your tracked website. To verify user activity, use Live Now.
After you complete these steps, the model training begins automatically within 24 hours. The training process completes and the model is ready to work before the start of the next business day.
Initially, you see only a green bar with a check next to it in the Outcome Scores section. This bar indicates that the outcome condition was met. You see outcome scores after your model is trained.

Ongoing training

The model training process is fully automated. You do not need a data scientist to start, monitor, or maintain the training. Your model is retrained nightly using the last 30 days of your user data. Also, your model is evaluated periodically and retrained on fresh data.  

While the model is undergoing retraining, outcome scores are predicted using the previously trained version of the model. During the training process, agents can continue working as normal.

Your newly trained model is tested to ensure that it performs better than the previous version in terms of its precision and recall. If the new model does not work at least as well as the previous model, the previous model is reinstated until more data is gathered. Historical models are not retained.
  • When you add a new outcome, your model training accounts for it automatically.
  • Automated model training and predictions are performed on the original customer journey events, which are not anonymized and generally contain PII.

Improve predictions

In general, the longer your models run and the more data they evaluate, the better their predictions are.

The best way to improve a model's predictions is to increase the number of achieved outcomes. In general, your dataset should contain several hundred positive examples for your model to be adequately trained to make reliable predictions.

Other factors can affect the accuracy of your model's performance, including:

  • Total number of visitor journeys recorded
  • Frequency that an outcome occurs in the data
  • Richness of events that the visitor produces
    You can exclude IP addresses to prevent events generated internally from influencing your models.

Predict based on custom events

You can provide predictions that other predictive systems generate as "custom events." The Outcome Score Service can consume the custom events and use them to train models and generate outcome predictions based on the features that an external system produces. This example shows how a generalized AI system enables hyper-personalization.

**We need to put this in plain English, but I want clarification before I begin...


Is this current-state, or tied to AI-23?