Customer Feature Matrix

A customer feature matrix is the structured customer-context representation used by contextual bandits, with one row per customer and columns for attributes, behaviours, and other predictive signals.

Key points

• Hightouch says the customer data warehouse already contains the information contextual bandits need; during decision-making, that data is structured into a customer feature matrix ^[src-026].
• Each row represents a customer, and each column represents an attribute or behavioural property such as purchase history, browsing behaviour, engagement patterns, demographics, lifecycle stage, preferences, days since last purchase, lifetime value, or email engagement ^[src-026].
• The contextual bandit uses this matrix plus possible actions and a machine-learning model to predict expected reward for each customer-action combination ^[src-026].
• Hightouch specifically names gradient-boosted decision trees as one model family useful in marketing because they can capture complex, layered patterns such as furniture interest after computer equipment purchases ^[src-026].
• The feature matrix is the bridge between a data warehouse and individual customer decisions; it converts raw customer data into usable context for Contextual Bandits ^[src-026].
• Braze gives the data-preparation version of the same pattern: raw events such as clicks or purchases need to be transformed into customer features such as recent clicks, recent in-app purchase value, engagement history, and contract state before AI decisioning agents can differentiate customers ^[src-152].
• Braze also notes that action features matter, not only customer features. An offer or plan can be represented by properties such as monthly price, upload/download speed, contract length, and discount terms ^[src-152].

Related entities

• Hightouch
• Braze

Related concepts

• Contextual Bandits
• AI Decisioning
• Treatment Personalisation
• Personalisation Gap
• Community Of Bandits

Source references

• ^[src-026] Hightouch — "Under the hood of AI Decisioning, part four: Contextual bandits"
• ^[src-152] George Khachatryan, Nathaniel Rounds, Victor Kostyuk / Braze – "A community of bandits" (source page: "From multivariate testing to AI decisioning", 2025-09-30)