How we've improved Arwen’s spam detection rate from 96% to 99%

Winning the social media spam arms race

Last week we released an innovative new spam detection model, which adds to our existing market-leading anti-spam capability. The new technique uses image analysis and sophisticated pattern matching to significantly increase detection rates. Arwen can now detect and remove 99% of spam comments, across all social networks, in real time.

Arwen’s existing spam classification feature has been incredibly successful at identifying and removing social media spam content. Arwen’s model is trained and optimised daily by its Data Science team, and also by customers flagging new spam patterns as they emerge. Both factors keep us on top of the problem.

However, fighting spammers is a cat and mouse game - as we raise new defences, so they work to circumvent them. AI has only increased the arms race, with spammers adopting AI to dramatically increase the rate at which they can innovate new spam patterns. We identified this trend early, and have been busy innovating their newest anti-spam model. 

Arwen’s new market-leading spam detection model brings two key innovations

• Arwen now runs sub-second image analysis of the avatar image of a commenting profile. When it detects common ‘spammer’ imagery in that profile - such as swimsuits, nudity, crypto images - it factors that into the real time moderation decision.
  
  As a result Arwen is now able to combine both comment classification and avatar image classification to make even more accurate decisions. For instance, in the example “edmonton” comment below, Arwen scores the comment text as only "somewhat likely" to be spam, however once in combination with the image analysis, it is able to make a far more confident "highly likely" 100% decision.

• Arwen now runs a more sophisticated pattern matching model on the text of comments, which is based on more complex detection rules than previous iterations. Not only does this greatly increase detection accuracy for spam comments, regardless of the length of the comment, but it also allows Arwen to quickly create new spam detection models as the spammers retaliate.
  
  For example, spammers have been experimenting with shortened single word comments for some time now, to get around traditional AI filters. Single words are notoriously hard for an AI to compute as they lack context. However, with Arwen’s latest model it is able to compare these comments to every spam comment it’s ever seen, to dramatically increase the accuracy of detection.

Leading the way in social media spam moderation 

These latest innovations add to Arwen’s unrivalled anti-spam capability. Arwen rests on a database of thousands of spam content patterns, built up over 4 years of processing over a billion comments globally. 

These are the hard yards that enable Arwen to achieve the tough incremental improvements that move detection rates from 96% to 99%.

All part of our mission to make our clients’ Social Media more social.