Applying Epsilon‑Greedy Bandit Algorithm for Content Delivery Optimization at DiDi

In the marketing growth domain, DiDi explored how engineering can empower business after the initial 0‑1 infrastructure is built. The focus is on quickly enabling business through data‑driven techniques rather than vague platform optimizations.

The team implemented the classic Epsilon‑Greedy algorithm from the multi‑armed bandit family and tightly integrated it with their Content Management System (CMS) for ad placement. The CMS controls the delivery of various assets (text, images, links, streams) across more than 600 slots, with competition among placements for limited resources.

Analysis revealed that many slots had low click‑through rates (CTR) and low utilization. To increase exposure value, the team designed a quality score based on the ratio of clicks to impressions, emphasizing placements with high exposure but low CTR. They applied min‑max normalization and a sigmoid function to smooth small differences.

Sampling was performed by pulling ~10% of traffic from Hive into MySQL, then evenly distributing this experimental traffic across all placements. A lightweight random‑number filter (1‑100) determines whether a request enters the sampling path. The sampling logic ensures uniform exposure across high‑dimensional conditions (city, audience, fence, etc.).

When a placement’s sampling quota is reached, the system falls back to the next eligible placement, allowing automatic intervention based on observed performance. Manual adjustments are also supported via a drag‑and‑drop UI, merging business‑defined priority with algorithmic ranking.

The optimization yielded dramatic results: after deploying the new service on a specific slot, the CTR rose from 1.35% on April 13 to 13.43% on April 20 (an 895% increase), and unique visitors increased from 211 to 1,659 (a 686% rise).

Additional engineering improvements include Redis hot‑key mitigation, batch memory operations, and careful handling of quality‑score scaling. The overall approach demonstrates that growth experiments can go beyond simple A/B testing by leveraging bandit algorithms and data‑driven sampling.