How Instagram Uses Artificial Intelligence and Machine Learning to Improve Instagram Content Management | Tech Reddy


Instagram is one of the most important sites on social media, and it is important to provide a personalized experience by providing the most important notifications to users. Instagram notifications are attached to the communication space between its users. Avoiding trivial things in notifications is very important to enhance the user experience. By determining quality in a big way, the engineers at Instagram found ways to send notifications that are more important and interesting to them.

A daily push notification for stories is a type of notification that lists shared content, ready to view, and content can be viewed directly while reading the post. In general, traditional Machine Learning models are used to predict the probability of clicking on an ad and determine the quality of the content in the ad. The lower the quality, the lower the chance of the ad being read to users. The engineers at Meta are applying Analytics and Machine Learning to improve the user experience and manage notifications every day. Also, using Causal inference and Machine Learning, the most active users have been identified, and the number of notifications sent to them has been reduced, improving the user experience.

It uses a traditional Machine Learning model called the Common Conversion Model (CTR) to predict the likelihood that a user will click on an ad. CTR models perform well in many applications across the industry. The predictions derived from these models have been used as a proxy to indicate user experience quality. If the predicted click probability is low, the notification will not be sent to the user and will be dropped in the middle because it is considered to be of low quality. CTR – the filter based on the template proved its worth as the average click rate for the CTR template was higher than without the template.

But by using CTR models, many ads were sent to users who are very active users of Instagram, and for these active users, there is no need to send these ads because they are already active them. they just see the news. This has put a task in front of the Engineers at Meta to send less notifications to active users as they will see the content even if they don’t receive notifications, and the painful thing is to identify these user. Because if active users are included due to the notifications sent to them on this list, they will not move again, and the User Experience and Engagement will decrease.

This leads to a user selection problem, and we want to increase efficiency by sending ads to the right groups. The solution to solve the problem is to use Machine Learning and Decision Making. To formulate the problem, we assume that there is a cost for the account to send each game ad each day, and a total budget for notifications to spend, so it is now a budget allocation problem. We will have to carefully check the increase in the delivery of the notice if it is not delivered. The increase in user capacity is defined as follows ui = Pri(active|do(send notification)) – Pri(active|do(drop notification)). For active user groups, the increment value is low, so sending notifications to these groups is not good, and users may reject it. User groups with high bounce values ​​can be selected by setting them down, and ads will be sent to them to improve user experience and manage ads. This also helps to improve value and budget. But again, how to determine the added value even before sending or dropping the ad? This is where causal inference problems and motivational modeling techniques come into play. Well, to apply a randomization sample, we design a randomized trial that allows for advertising or reduction, and collects data from this randomized trial. Based on the data collected, a network-based dynamic model was created to predict the increase in interest between posting and not posting a daily game post, then leaving the solution to the limited budget allocation problem. But in reality, notifications are created and scored online. Therefore we cannot determine the inflation effect for all notifications in advance. So there needs to be an online way to decide which ad to post and which one to post. A simple solution to this is to have a fixed threshold and compare it to the generated score, and if the score is above the threshold value, the ad is sent, resulting in a constant ad delivery rate r (0

Using this model, users and notifications with a high impact of the increase will be targeted, resulting in a decrease in delivery volume compared to the CTR model without a decrease in user engagement and additional benefit of reduction in resource use.

Reference Articles

Avanthy Yeluri is a PhD student at IIT Kharagpur. He is very interested in Data Science because of its many applications across various industries, as well as its technological advances and how it works in everyday life.


Source link