Ad frequency control (frequency capping)

Frequency Control Introduction

Frequency Control in ads refers to the number of times a user sees an ad (or similar AD) for up to a specified period of time, such as advertisers can limit a user to a maximum of one ad per day 3 times (frequency control also allows Publisher to specify, but this article does not consider publisher).

The frequency control mentioned in this paper is not rigid, that is, the number of impressions will only reduce the probability of the same ad appearing, but not to achieve a certain number of times after the total. The difference between the two is: the former advertisers need to try to find out how long the time of the most appropriate, more trouble, but if he gets close to the number of facts, will be beneficial to him. The latter is the convenience of advertisers, the price is that the advertising platform will only optimize the ECPM, not to consider the interests of advertisers.

The introduction of frequency control has the following advantages:

1. increase the number of reach audiences. Frequency control allows more audiences to see ads. Because frequency control eliminates the advertiser's limited budget, it is always consumed by the constant display of ads to the same group of people.

2. increase CTR. Frequency control is one of the effective means to improve the ad CTR . CTR is low after about 4 impressions of the same ad , as users have already begun to ignore the ad.

3. improve CVR(conversion rate), similar to CTR , at the first show the conversion rate is the highest, after the show 4-6 times basically no conversion.

Frequency Control principle

See, the horizontal axis is the number of impressions, the ordinate is CPM, the curve is different ads (yes, I know that the picture and I said is not the same thing, but the truth and data is really similar, because I can not take the company's data). It is important to note that:

1. all ads are displayed as the number of impressions increases and theCPM decreases.

2. take the Blue Line and red line for example, the blue ad shows 1 times the CPM is higher than the red display once the CPM, but the blue shows two times after the CPM is a CPM that is shown below the red one .

3. There is a difference in the speed at which each polyline falls.

Recall the formula:CPM = CTR * Bid,ECPM = pctr * Bid. The Bid in the formula is known and fixed, so that is ECPM because the impressions change because the impressions affect the pctr, then if the impression is based on the number of impressions PCTR has been adjusted, then the ECPM is more reasonable.

Considering the number of times it has been shown to adjust the pctr , ideally we will reap a higher CPM, which means the company can make more money. Can you imagine the way I feel when I make this remark? )

Frequency control after the figure should be so, the Purple Line is optimized after the effect, if you do not understand, the leftmost is actually less a black vertical bar (I emphasize again, the figure of ad Network to think of it as advertising, if you are SSP, then don't think about it). You can see that the Purple Line is significantly higher than the blue Line.

Frequency Control implementation

first of all, based on the impressions of the pctr to adjust whether it is pctr own problems, this I have been a little doubt, the number of impressions can obviously be added as a feature to pctr 's model, but I haven't tried it myself. However, whether or not impressions are included as a feature, the approach is similar, and it is important to get the data.

           data in two parts, part of the real-time data, and another part of the offline data. Need storm collect real-time data because write hive database typically has a time delay of the hour. kv DB is the result of aggregating real-time data and offline data, < Span style= "FONT-FAMILY:CALIBRI;" >key is user , value is a list of

If it is the number of impressions as the characteristics of pctr Learning, then all OK , if not, then trouble, first of all, to get the line in front of the chart, to get impressions of each ad the impact of PCTR. Here is a problem, advertising exposure to do? Click on dozens of, how to count? Is the dimension up to calculate the promotion plan? What if the exposure is not enough? Then up to the ad category? The main dimension of the advertiser? Or more directly, pretending not to see the data, directly ignored, the goal is only to solve the statistical significance of advertising.

Think of a little more questions, the crowd a may have seen once not point will not point, the population B has seen the probability of a click and has never seen the click probability of the same. The definition of a crowd can be infinitely diverse. Another ad was shown at the top ad bit, and the bottom ad bit, do you think it was shown once? So I haven't tried it, but I think it's easiest and reasonable to put the impressions in the PCTR model as a feature.

Partial internal citation:

http://www.masternewmedia.org/online-advertising-management-frequency-capping-to-optimize-ad-revenues/

Ad frequency control (frequency capping)