TechPrufer
Get started
Slack
Slack logoSlack · Feature trail

Machine Learning for Invite Spam Detection

2 posts · Jun 2017 – Aug 2020

What this trail covers

This thread tracks the evolution of Slack's systems for combating invite spam. Initial efforts focused on manual rule-based systems and rate limiting. This post details the development and deployment of a machine learning model to proactively identify and block spam invites, significantly reducing false positives and freeing up engineering time. The ML approach leverages a sparse logistic regression model trained on features derived from invite data, aiming to predict invite acceptance rates as a proxy for spam.

All Slack trails
  1. Blocking Slack Invite Spam With Machine LearningThis post details the implementation of a machine learning model to block invite spam. It describes the transition from a labor-intensive, hand-tuned rule-based system to a supervised ML problem. The approach uses invite acceptance rate as a proxy label and trains a sparse logistic regression model on approximately 60 million features, including user/team IDs, email addresses, domains, IP addresses, text stems, character sequences, websites, and team age. The post also covers model serving via an in-house Kubernetes-based microservice and highlights the impact of reduced false positives and automated human effort.
  2. Introducing HighlightsThis post details the technical implementation of Slack's 'Highlights' feature, which uses machine intelligence to predict important messages. It describes the 'work graph' concept for capturing communication relationships and the use of logistic regression models with over a hundred features to predict user engagement. The post also addresses class imbalance issues through stratified sampling and prior correction, and discusses personalization by grouping user behaviors and incorporating user-specific features from the work graph. A justification layer is also described, which uses feature weights to explain why a message is highlighted.