Contextual Bandit
A multi-armed bandit that conditions arm choice on observable context features (time, audience, device).
Last updated
Definition
Vanilla MAB picks the same arm regardless of who's looking. Contextual MAB conditions on context: 'the best creative for a Mumbai user on iOS at 8 PM might differ from a Bangalore user on Android at 10 AM.' Used by journey-orchestration systems to personalize variant selection per audience cohort.
How it applies in India
Because Indian customers chat in different languages, the customer’s language is a natural context: WatEase’s Copy Optimizer keeps separate evidence for each segment, such as a language.
An example
A business tests two versions of a festive message, one in English and one in Hinglish. Across all customers the English version wins narrowly, but among customers who chat in Hindi the Hinglish version wins clearly. A plain bandit would end up sending English to everyone; a contextual bandit learns that the answer depends on the segment and sends each group its own winner.
How WatEase uses it
WatEase uses a contextual bandit in its Copy Optimizer. Each message variant keeps separate evidence for every customer segment and for every goal you measure, and a weighted draw picks the variant to use; a segment with no evidence yet borrows the variant’s overall record. The context is a segment label, such as the customer’s language, rather than a full profile. Auto-Tune shows the traffic split the draws produce, and customers in the holdout always receive the oldest variant so they never count as evidence. It is one of the ways agentic marketing learns without a person reading every report.
Frequently asked questions
Where does WatEase use a contextual bandit?
In the Copy Optimizer, which tests message variants per customer segment and per goal. The variant for each segment is chosen by a Thompson-sampling draw over that segment’s own results, falling back to the variant’s overall results while the segment is new.
Related terms
- Thompson SamplingA Bayesian way to split traffic between options — creatives, templates, audiences — that sends most of it to the likely best one while still testing the rest.
- Multi-Armed BanditA class of algorithms (Thompson, UCB, ε-greedy) that balance exploring uncertain options against exploiting known winners.
Try WatEase
Run your WhatsApp commerce on the platform built for India — Cloud API, GST invoices, UPI checkout, opt-in tracking, and 0% markup on WhatsApp messages (competitors add 20–25%). Pricing is on request — quoted by the WatEase team.
Get Started Today