[ad_1]
Instagram changes the way to create ads from CTR to factoring with ML models to identify the most active users who are likely to receive ads.
CTR (click-through rate) is one of the most used algorithms to predict the content’s effectiveness for the user based on the probability of clicking on the ad. If there is a very low probability that the notification will fail in the delivery stream and the user will not receive the notification. CTR-based analysis works well for daily game notifications, however, a large portion of notifications are sent to users who are actively using Instagram. The goal is to provide better notifications to users and to accurately identify these types of users, who seem to be active or active when receiving a notification, to reduce the risk of reduced participation by users. user.
In this way, the problem becomes a user selection problem: increasing the efficiency of advertising by choosing the right users. The solution adopted is a combination of reasoning and machine learning.
Considering the fixed cost of posting an ad and the daily budget to post these ads, it becomes a budget allocation problem. The solution to the problem is to calculate the added value to send the message instead of sending it. It can be modeled mathematically as follows:
\(ui=Pri(action|action(send notification)) – Pri(action|action(drop notification))\)
For power users, sending notifications will be ineffective and may cause rejection. To optimize the menu, the message can be sorted by value in descending order, and only the top notifications with the highest value will be sent. In this way all added values are maximized with a limited delivery budget.
Estimating the increase in value of a pre-release ad is a difficult problem. It is a modeling problem, and it is possible to use convolutional modeling techniques to solve it. To use the biased model, the Meta engineers created a randomized experiment in which each ad was randomly assigned or not. Based on this data a neural network based on the activation model is developed to predict the increase in the value of the posted ad.
The score generated online is compared to a set threshold to determine whether an ad will be sent or dropped, but from the data collected it can be seen that the difference in the sending rate is due to the increase determined by the ML models . Meta engines use an online metric calculation service to adjust the estimates to a uniform distribution while maintaining orders, to ensure a consistent delivery rate.
Decision flow for advertising management
Using this model and focusing on users improves the user experience, reduces resource usage and does not decrease user engagement.
[ad_2]
Source link