User feature matrix

Overview

The user feature matrix (UFM) is a structured set of customer attributes and behavioral features used by AI Decisioning (AID) to evaluate what is most likely to drive a desired outcome for each individual user.

The UFM is built from data in your data warehouse and made available to AID through your Hightouch schema.

Fundamental concepts

• Features are customer attributes and behaviors that can help predict which actions drive your goals.
• High-signal features typically span behavioral history, customer value, demographics, preferences, and current context.
• More features are better. AID will learn which ones matter most.
• Start with the data you have and enrich over time.
• The feature matrix stays in your data warehouse (secure, governed, and under your control).

Why the UFM?

Traditional marketing treats customers in broad segments—"high-value customers," "recent purchasers," or "email engagers." AI Decisioning moves beyond segments to truly individualized experiences.

How AID uses the UFM

AI Decisioning uses UFM features to make decisions at the individual user level. Instead of treating users as part of a single segment, AID evaluates each person based on their unique combination of attributes and behaviors.

Specifically, AID uses UFM features to:

• Understand individual context by considering how multiple features interact (for example, recent behavior, lifecycle stage, and preferences).
• Identify patterns that predict outcomes by learning which feature combinations are most likely to drive goals like clicks, purchases, or conversions.
• Continuously improve decisions as new outcomes are observed and incorporated into future predictions.
• Scale personalization automatically by making individualized decisions for millions of users without manual segmentation or rules.

Who builds the UFM?

The UFM is custom built by Hightouch AI Decisioning Machine Learning Engineers (MLE) as a related model in your Hightouch Schema.

How to gather data for the UFM?

The UFM features are scoped prior to launching AID Agents with the AID Hightouch team, including your dedicated Machine Learning Engineer (MLE). The relevant data tables must be surfaced in your Schema by your data team. It is recommended to leverage the UFM tab in your Engagement Workplan to document details about each data/features.

Categories of High-Signal User Features

The most effective feature matrices typically include features across these key dimensions:

1. Behavioral History Features

These capture what customers have done and how they interact with your brand and products.

Examples:

• Purchase history and recency (days_since_last_purchase, total_purchases_90_days)
• Engagement patterns (most_engaged_day_of_week, email_open_rate_30_days, app_sessions_last_week, last_date_email_click)
• Product interaction (product_categories_viewed, cart_abandonment_count)
• Support interactions (support_tickets_90_days, avg_ticket_resolution_time)
• Conversion Mix/top (most_purchased_category,last_date_order,Purchase_day_of_week, purchase_time_of_day)

Why they're high-signal: Past behavior is typically the strongest predictor of future behavior. These features reveal customer preferences and intent.

2. Demographic & Firmographic Features

These provide stable contextual information about who the customer is.

Examples:

• Location data (country,state,timezone, urban_vs_rural)
• Demographic attributes (age_range, gender,account_type)
• Device preferences (primary_device_type, operating_system)

Why they're high-signal: While not as predictive alone, these features interact powerfully with behavioral data to create nuanced segments.

3. Customer Value & Lifecycle Features

These indicate where customers are in their journey and their economic relationship with your brand and products.

Examples:

• Lifetime value (total_revenue, avg_order_value)
• Customer tenure (months_as_customer, days_since_signup)
• Subscription status (subscription_tier, months_until_renewal, in_last_month_of_contract, signup_date,)
• Churn risk indicators (engagement_decline_30_days, product_usage_frequency)

Why they're high-signal: These features help AI understand customer loyalty, price sensitivity, and which interventions are appropriate at each lifecycle stage.

4. Preference & Engagement Features

These capture explicitly stated preferences and communication response patterns.

Examples:

• Overall engagement (emails_click,push_click,email_response_%, push_response_%)
• Channel preferences (preferred_contact_method, email_vs_sms_engagement)
• Content preferences (favorite_product_categories, preferred_brands)
• Promotional sensitivity (discount_redemption_rate, responds_to_urgency)

Why they're high-signal: These directly inform which messages and channels will resonate with each individual.

5. Current Context Features

These capture the customer's immediate situation and needs.

Examples:

• Product ownership (current_plan_speed,devices_owned)
• Contract status (contract_end_date,eligible_for_upgrade)
• Recent activity flags (browsed_last_24hrs,recent_search_terms)
• Account health (payment_method_expiring,billing_issues)

Why they're high-signal: These features enable timely, contextually relevant outreach that meets customers where they are right now.

Feature Engineering Best Practices

Start with What You Have

You don't need perfect data to begin. Our AI Decisioning team can work with the customer features you already have in your data warehouse. Start with:

• Basic behavioral events (purchases, logins, page views)
• Demographic information from your CRM
• Engagement history from your marketing platforms

Turn Raw Events into Features

Raw event logs need to be aggregated into meaningful features. Our MLE can assist you with these aggregations. For example:

Raw event: email_click events with timestamps
Engineered features:

• email_clicks_last_7_days
• email_clicks_last_30_days
• days_since_last_email_click
• email_engagement_trend (increasing/decreasing)

More Features Are Better Than Fewer

The AI models can handle hundreds of features and will automatically determine which ones matter most. In general, it's better to include a feature that might be relevant than to exclude it.

If a particular feature ends up not doing much to predict customer behavior, the model will learn to ignore it.

Getting Started: Feature Discovery Workshop

To help you think through which features to provide, consider these questions:

1. What distinguishes your best customers from others?

• These differentiators should be captured as features

2. What information do your best marketers use when personalizing manually?

• These instincts can often be quantified as features

3. What customer attributes correlate with different outcomes?

• Purchase behavior, engagement, churn, upsell success

4. What contextual information influences customer decision-making?

• Contract status, product ownership, recent interactions

5. What data lives in your warehouse but hasn't been used for personalization?

• Support tickets, product usage logs, browsing behavior

Next Steps: Sharing Your Feature Matrix

We’ll need to surface your data in the Schema as related models and reference details about the features and tables in the UFM tab of your Workplan Engagement plan

1. What customer data exists in your data warehouse?
   • Tables, key attributes, event streams
2. Which features are already calculated?
   • Existing aggregations, scores, or derived attributes
3. Which features need to be engineered?
   • Hightouch can help transform raw events into features
4. What business context should inform feature selection?
   • Your unique understanding of what drives customer behavior

The goal is to leverage the data you already have while identifying opportunities to enrich your feature matrix over time. AI Decisioning works with your existing data infrastructure. There is no need to move data or rebuild systems.

Example: E-Commerce Feature Matrix

Here's what a practical feature matrix might look like for a retail brand:

E-Commerce Feature Matrix