Katkat
Back

Evaluate fake accounts and ad creation

MetaData Scientist · Onsite
MediumData ScientistOnsiteAnalytics & Experimentation

Last updated

This question tests the ability to design an evaluation framework for identifying malicious actors in an ad-tech environment and measuring the performance of automated content generation tools.

Scenario

You are a Data Scientist at a large social media platform. Two critical initiatives are currently under review:

  1. Fake Account Detection: You have a new machine learning classifier designed to identify and suspend 'inauthentic' accounts. The system currently triggers an automatic flag if the model score exceeds 0.85.
  2. AI Ad Creation: A new tool allows advertisers to automatically generate ad creative using generative AI. There is a concern that this increases the volume of low-quality or policy-violating ads entering the auction.

Task

  1. Metric Design: Propose a primary metric to evaluate the efficacy of the fake account detection system. Explain why simple accuracy is insufficient and how you would account for the 'unlabeled' population of accounts that never trigger the system.
  2. Experimentation: Suppose you want to test the impact of the AI Ad Creation tool on long-term user retention. How would you design a randomized controlled trial (RCT) to measure this, considering potential network effects where the experience of one user might be influenced by ads shown to another?
  3. Trade-offs: If the fake account detector model is adjusted to increase precision (to reduce false positives), what are the potential business impacts on the ad ecosystem, and how would you quantify the 'cost' of a false positive versus a false negative?

Hints

Hint 1

Consider using a hold-out set of accounts that are manually audited by human reviewers regardless of their model score to estimate the true prevalence of fake accounts.

Hint 2

For the ad creation tool, think about using a 'cluster-level' randomization or a 'switchback' experiment if individual-level randomization is not possible due to network interference.

Hint 3

Frame the trade-off in terms of the cost of losing a legitimate advertiser (False Positive) versus the cost of poor user experience and potential churn due to malicious ad exposure (False Negative).

Follow-up questions

  • How would you identify if the fake account detection model is introducing systematic bias against specific demographics or geographic regions?
  • If the AI-generated ads result in a short-term increase in click-through rate (CTR) but a decrease in 30-day user retention, how do you reconcile these conflicting signals?
  • What statistical methods would you use to account for 'selection bias' if advertisers are allowed to opt-in to the AI generation tool rather than being randomly assigned?