How Private History Can Supercharge E‑commerce Recommendations: The PH‑MAB Mechanism Explained

Public and Private History in MAB Models

Multi‑armed Bandit (MAB) models are a key subfield of artificial intelligence and reinforcement learning, widely used to describe e‑commerce recommendation problems. In a classic MAB setting, a slot machine has multiple arms, each yielding stochastic rewards; the player (the platform) selects arms over N rounds to maximize total reward.

The paper Multi‑armed Bandit Mechanism With Private History proposes a new mechanism, PH‑MAB, that leverages both public history (CH) – information about a product’s recommendation outcome observed by both platform and seller – and private history (PH) – additional transaction data sellers have from other channels. By encouraging sellers to report their PH truthfully, the platform can incorporate richer information into the MAB decision process.

To ensure truthful reporting, the authors design a mechanism rooted in mechanism design theory that makes reporting real PH a dominant strategy for rational sellers. The mechanism rewards sellers based on reported histories and observed rewards, and it can be proved that truthful reporting yields the highest expected payoff for each seller.

Experimental simulations demonstrate that PH‑MAB achieves lower regret than the standard epsilon‑greedy algorithm and yields higher platform revenue, confirming the benefit of integrating private histories.

The approach also highlights that many real‑world multi‑agent problems—such as clinical trial arm selection or routing decisions—can be modeled as MAB problems, suggesting broader applicability of the PH‑MAB design.