Facebook Inc.
Deep Entity Classification, or DEC, is Facebook's machine learning model that, over the last year, has helped Facebook delete or prevent the creation of 6.5 billion accounts. DEC has been particularly helpful in the latter method of catching fake accounts before they can be created.
When scoping out live accounts, DEC can study an account's behavior and determine if it is either a regular user or a fake account attempting to emulate normal Facebook behavior (such as friending other users) while pushing disinformation or trying to scam other Facebook users. According to Facebook, DEC analyzes around 20,000 profile traits to conclude whether an account is fake or legitimate. DEC's efficiency and vast scope has helped Facebook to reduce the volume of fake accounts by 27%. DEC is also self-teaching, enabling it to discover new methods used by malicious users and adapt.
Before DEC, Facebook was mostly reliant on users reporting suspect accounts. Due to how slow the reporting process was, many accounts would linger for some time before being reported, due to how convincing some can be to average users. An unintended side-effect of this process was the mental toll it took on Facebook moderators, who were often contractors unaffiliated with the company itself. Contractors were exposed to a broad array of horrific while tracking user reports, and often had little in the way of support for the toll they suffered.
Facebook is no stranger to disinformation and fake accounts, as it has been the target of intense scrutiny for its part in disinformation campaigns during the 2016 elections and the rapid proliferation of fake news stories over the last few years. As the U.S. approached the 2020 election, Facebook found itself under even more scrutiny and was flooded with calls to regulate its content better. Facebook's development and deployment of DEC seems to be an earnest attempt by the company to address these problems, streamlining the process of combating disinformation and scamming while increasing efficiency.
