https://blog.csdn.net/LW_GHY/article/details/71455535
ADX bid adjustments,
After the reduction of estimated ctr sampling adjustment
2. dynamically adjust our offer
in the offer links DSP, the click-through rate for each traffic prediction model would forecast a CTR, adx to meet the needs of the market, will add items index and the coefficient adjustment. But this offer by way traffic acquisition, due to changes in external competitive environment, the natural flow at different time intervals, often unstable CPC will appear. Coefficients of the offer are the same for all the media, and the general quality of the media are all reserve price, the reserve price and inconsistent different media, resulting in the offer can not be applied to all manner of media, some high-quality media can not get enough traffic appears .
Our goal is to under certain circumstances CPC, in high-quality media, quality time as much as possible to get traffic, where we need to adjust our offer dynamic line based on real-time feedback and expectations of stable CPC [10]. So in the auction environment, time of day, when the media change, CPC remains stable, further ensures our revenue maximization (the same marketing costs, get the most traffic).
4. Sampling negative samples
due to the extremely uneven distribution of positive and negative samples ad clicks outside the station CTR is generally low, the vast majority of the sample is negative samples, in order to ensure that the model recall positive samples, the need for negative samples according to a certain proportional sampling.
5. Real CTR calibration
due to the negative sample after sample, it will create the illusion of high click-through rate, you need to restore the predictive value of the true value. The adjustment formula is as follows:
q: The actual click-through rate adjusted.
p: Negative Samples Sampling estimated at CTR.
w: Negative sample sampling ratio.