[Legacy]{class="badge informative"}

Personalized optimization model personalized-optimization-model

By leveraging the state-of-the-art technologies in supervised machine learning and deep learning, Personalized optimization allows a business user (marketer) to define business goals and utilize their customer data to train business-oriented models to serve personalized offers and maximize KPIs.

Unlike non-personalized ranking which optimizes based on the global performance of each offer, Personalized optimization learns the relationship between an individual customer’s attributes and the offers most likely to drive the chosen KPI for that customer. The result is an offer selection tailored to each profile rather than a single best offer served to everyone.

Use cases and benefits use-cases

Personalized optimization is well suited to decisioning scenarios where different customers respond differently to the available offers and where the catalog of offers is meaningfully differentiated and doesn’t change often. Common use cases include:

Key benefits:

Dataset requirements dataset

To train a personalized optimization model, the dataset must have at least two offers with at least 250 display events (for example, impressions) and one success event (for example, click or conversion) within the last 30 days.

Offers with fewer than 250 display events and/or no success events within the last 30 days will remain eligible for inclusion in the exploration traffic. They will also be eligible for inclusion in the personalization traffic but will be treated as equivalent to the worst scoring predicted offer in decisioning, until they meet the required minimum display/success events and the model gets re-trained.

Until the first time a personalized optimization model is trained, offers within a selection strategy utilizing a personalized optimization model will be served at random.

How it works how

The model learns complex feature interactions between offers, users’ information and contextual information to recommend personalized offers to end users. Features are inputs into the model.

There are 3 types of features:

The model has two phases:

The model supports optimization of continuous variables (such as revenue and customer lifetime value) in addition to binary variables (such as clicks and conversions). Predicted values for a binary metric such as clicks will always be between 0 and 1. Predicted values for a continuous metric such as order value will always be a number greater than or equal to zero. Ranking scores are normalized to ensure consistent behavior across both metric types when used in formulas or comparisons.

Illustrative example illustrative-example

Binary response (conversion) binary-response

Consider a simplified dataset of historical interactions between users and offers. Each row records an offer that was shown, two customer signals — loyalty tier (high = 1) and whether the customer opened a recent email (yes = 1) — and whether the customer converted (yes = 1).

For Offer A, conversion is more likely when both signals agree (both high or both low). For Offer B, conversion is more likely when the email was opened, regardless of loyalty tier. Based on the learned pattern, the model can predict the better offer for each customer based on their signals.

Figure 1: In the highlighted mismatch row, Offer A was shown when the signals disagreed and did not convert. Based on the learned pattern, Offer B would be the better recommendation for that customer next time.

This is the essence of the approach: learning and memorizing historical feature interactions and applying them to generate personalized predictions for each customer.

Continuous response (revenue) continuous-response

The same idea extends to continuous outcomes. Instead of predicting whether a customer converts, the model predicts a continuous value (expected revenue) for each offer and customer segment, and ranks offers by that predicted value.

Figure 2: Predicted revenue for two offers across four customer segments. For high-loyalty customers who opened the email, Offer A is expected to drive the most revenue; for low-loyalty customers who opened the email, Offer B is the stronger choice. The model selects the offer with the highest predicted value for each segment rather than applying one rule to all customers.

Ensemble model components ensemble

Personalized optimization is delivered as an ensemble model — several complementary model arms run together, and a supervisory layer decides how much live traffic each arm receives. This design lets the system pursue two goals at once: learning which offers perform best (exploration) and serving the offers already known to perform well (exploitation).

Balancing exploration and exploitation

Every decisioning system faces a tradeoff between exploring under-tested offers to gather information and exploiting proven offers to maximize immediate return. Reserving too little traffic for exploration leaves high-potential offers undiscovered; reserving too much sacrifices lift on offers that are already performing. The ensemble manages this balance automatically by holding a minimum exploration floor while shifting the remaining traffic toward the better-performing personalized arms over time.

The ensemble is composed of four traffic arms:

Uniform random (exploration arm) uniform-random

The uniform random arm assigns offers to customers at random from among the eligible offers. Because it does not favor any offer, it produces unbiased data about how customers respond across the full catalog — the raw material the personalized arms learn from. It is the only arm active before the first model is trained, and afterward it continues to hold a minimum exploration floor so the system keeps learning.

Neural network (personalized arm) neural-network

The neural network is a personalized arm that predicts the best offer for a given customer based on their attributes and audience memberships. It learns complex, nonlinear interactions among offers, customer features, and context, and is well suited to capturing subtle patterns across many features.

Contextual bandit (personalized arm) contextual-bandit

The contextual bandit is a second personalized arm that also predicts the best offer for each customer based on their audience memberships, using a bandit approach that continually balances learning and performance as it serves. Running it alongside the neural network lets the ensemble draw on the strengths of two distinct personalized methods.

New offer booster (non-personalized arm) new-offer-booster

The new offer booster is an overall-winner Thompson Sampling bandit (non-personalized) that makes optimistic assumptions about the performance of new offers — those with few recorded impression events within the model lookback period. This gives promising new offers the early exposure they need to prove themselves, addressing a known cold-start shortcoming in which the model otherwise struggled to direct enough traffic to new or high-performing but restrictively eligible offers.

How traffic is allocated across the arms traffic-allocation

At initialization, no model has trained yet, so 100% of traffic goes to the uniform random baseline — the only arm with a learned distribution to sample from. After the first successful training run, each arm receives a minimum traffic floor (5%), and the supervisory bandit allocates the remaining traffic based on observed performance. As the model trains across successive rounds, traffic converges toward the highest-performing arms with a maximum possible allocation of 85% traffic.

Figure 3: A possible traffic allocation trajectory across the four ensemble arms at initialization and across successive training rounds. At initialization, all traffic flows to the random baseline. After each training run, the supervisory Thompson Sampling bandit shifts allocation toward better-performing arms, while maintaining minimum 5% traffic. Actual allocation will vary based on observed arm performance.

Key model assumptions and limitations key

In order to maximize the advantage of using personalized optimization, there are some key assumptions and limitations to be aware of.

Cold-start Problem cold-start

Cold-start problems occur when there isn’t enough data to make recommendations. For personalized optimization, there are four types of cold-start problems.

Re-training re-training

Models will be re-trained to learn latest feature interactions and mitigate model performance degradation weekly.