Instagram has shared new details on how its application uses machine learning technology to illustrate content to users. The company said that when making recommendations, technology focuses on creating accounts that people think they will enjoy, rather than individual posts.
According to the platform, more than half of its 1 billion users visit the Explore tab to discover videos and photos each month.
Building the core recommendation engine, which manages billions of content uploaded to Instagram, is an engineering challenge.
The details of the post are technical in nature, but they offer an interesting behind-the-scenes perspective at a time when algorithm recommendation systems are being scrutinized for pushing users towards dangerous, obnoxious, and extreme content.
While not criticizing Instagram as fiercely as YouTube, it has its share of problems. Hateful content and misleading information on the platform are growing like any other social network.
Certain mechanisms in the application, such as the suggestion follow-up feature, were highlighted and talked about driving users towards extreme views of topics such as anti-vaccination.
Instagram engineers explain the operation of the Explore tab while avoiding thorny political problems.
Ivan Medvedev, software engineer at Instagram, said: This is the first time that we have entered with very specific details about the foundational building block that helps us to provide customized content on a large scale.
The post emphasizes that Instagram is huge and the content is very diverse, with topics ranging from Arabic calligraphy to train models.
This is a challenge to recommend content that the platform is trying to overcome by focusing on accounts that might be of interest to users rather than on posts that users might want to see.
Instagram identifies similar accounts with each other by adapting a common machine learning method known as word embedding. Word embedding systems examine the order in which words appear in the text to measure their relevance, and the platform uses a similar method to determine how closely two accounts relate to each other.
To make its recommendations, the Explore system starts by searching the primary accounts, accounts with which users have interacted in the past by liking or saving their content.
The system identifies accounts similar to those of which 500 content is selected, filtered content is filtered to remove spam and misleading content, as well as content likely to violate the platform policy.
The remaining posts are categorized based on how likely the user interacts with each one, and the top 25 posts are sent to the first page of the user's Explore tab.
One of the most exciting parts of the “Explore” tab is the ongoing challenge of finding new and interesting ways to help the platform community discover the most interesting and relevant content on Instagram.
Comments
Post a Comment