Problem
A supply-side platform (SSP) sells ad slots through real-time auctions and receives up to 25 million auction requests per hour. Only 15-20% of the inventory is high quality, since bots inflate the auction count on junk sites. Even on good sites, at most 5% of requests get a positive response from advertisers' platforms (DSPs). Requests without a response waste hardware load and data transfer.
Example from the article: at Teads, a model that skipped requests unlikely to get a bid cut SSP infrastructure cost by 13% in header bidding and 10% in classic integrations, with no loss of ad revenue.
Solution in brief
Put a small classifier right after the request is enriched and before bid requests go out. It predicts the probability of a useful outcome and stops the auction, or skips a specific DSP, when the probability is below a threshold. The threshold is computed dynamically to hit a target criterion. The scheme fits any expensive downstream call with a low hit rate.
Questions this case answers
- How to skip expensive calls to external partners that rarely return a useful answer by predicting the outcome before the call, within a few milliseconds per request
- How to set a decision threshold dynamically so that a target cost or margin criterion holds as traffic changes
- How to cut infrastructure costs by filtering low-probability requests before expensive calls
- How to decide which bid requests an SSP should send to which DSPs
Tools and methods
- Publisher inventory filter — predicts whether any buyer will bid on a slot for a given enriched request
- DSP throttling model — predicts whether a specific DSP returns a non-null bid in time
- Deep & Wide network — compact model with inference under 5 ms
- Dynamic threshold — set per publisher to meet a given criterion
- A/B test on programmatic margin — measures the effect on the DSP side
Google or a one-time email code. No password.
