Anomaly detection with Variational Autoencoders
Submitted by Aditya Prasad Narisetty (@adityaprasadn) on Saturday, 14 April 2018
There are two types of companies: those that have been hacked, and those who don’t know they have been hacked - John Chambers
Similarly, a firm’s promotional marketing(new user signup, referral, loyalty and cashback) is vulnerable towards user exploits. This abuse can run into millions of dollars in short notice before you wake up to it.
Also, malicious user activity on your platform can result in bad user experience. Especially on
1. Multi-agent game playing platforms
2. E-commerce with fake reviews
In this talk we’ll look at a novel approach using Variational Autoencoders to potentially identify a fraudulent transaction.
- Abuse patterns on E-commerce platforms
- Rule engines as gate keepers in transactional systems
- Supervised Machine Learning models and their effectiveness
- Limitations of Supervised M/L models
- Reconstruction error as a measure of anamoly.
- Introduction to Autoencoders and intuitive understanding of Variational Autoencoders (VAE)
- Benifits and limitations(what they can/can’t do) of using VAE.
- VAEs in action for detecting fake user registrations @Dream11.
Brief understanding of Anamoly Detection and enthusiasm to learn :)
Aditya Narisetty is a regular speaker in the machine learning community with 5+ years of experience in M/L and D/L. He’s a Sr Data Scientist @Drem11, India’s biggest fantasy sports platform with 2.5Cr+ users. He loves to share his work with the community and spoke at events like Google DevFest, Anthill Inside, Bootcamp @IIT-B, DataNativesX and Mumbai AI Meetups. He’s an IIT-B graduate and Ironman 70.3 Triathlete.