Measure fake-news interventions under network interference

Quick Overview

This question evaluates a data scientist's competency in causal inference and experiment design under network interference, covering cluster-randomized trials, exposure mapping, estimation of total, direct and spillover effects, power analysis, robustness checks, and detection of adversarial adaptation.

Measure fake-news interventions under network interference

Company: Meta

Role: Data Scientist

Category: Analytics & Experimentation

Difficulty: hard

Interview Round: Onsite

You test a label that warns users before resharing a suspected fake-news link. Because exposure and behavior propagate on the graph, SUTVA is violated. Design an experiment that identifies total, direct, and spillover effects using cluster randomization: (a) how you construct clusters (e.g., graph partitioning to minimize cross-cluster edges), (b) exposure mapping to define treated vs indirectly-exposed users, (c) primary outcomes (reshare rate, unique reach) and misinfo prevalence reduction, (d) estimation strategy (Horvitz–Thompson or difference-in-means with cluster weights) and power implications of cluster size, and (e) sensitivity analyses for contamination. How will you detect strategic adversarial adaptation during the test?

Quick Answer: This question evaluates a data scientist's competency in causal inference and experiment design under network interference, covering cluster-randomized trials, exposure mapping, estimation of total, direct and spillover effects, power analysis, robustness checks, and detection of adversarial adaptation.

|Home/Analytics & Experimentation/Meta
Meta logo
Meta
Oct 13, 2025, 9:49 PM
hardData ScientistOnsiteAnalytics & Experimentation
5
0

Experiment Design Under Interference: Warning Label for Suspected Fake-News Reshares

Context

You are testing a pre-reshare warning label for links suspected to be misinformation. Because both exposure and behavior propagate along the social graph (e.g., seeing friends' reshares, feed ranking, social reinforcement), the Stable Unit Treatment Value Assumption (SUTVA) is violated: one user's treatment can affect others.

Design a cluster-randomized experiment that can identify total, direct, and spillover (indirect) effects despite interference.

Task

Describe a design that includes:

(a) Cluster construction

  • How will you partition the graph into clusters to minimize cross-cluster interference? Outline practical constraints (e.g., cluster size bounds) and algorithms.

(b) Exposure mapping

  • Define what counts as a treated user vs an indirectly-exposed user under partial interference. State any thresholds or functions of neighbors' treatment you will use.

(c) Outcomes

  • Specify primary outcomes, such as reshare rate and unique reach. Include a system-level metric for misinformation prevalence reduction.

(d) Estimation strategy and power

  • Describe estimators for total, direct, and spillover effects (e.g., Horvitz–Thompson or difference-in-means with cluster weights). Discuss how cluster size and intra-cluster correlation influence power.

(e) Sensitivity analyses

  • Propose checks for contamination (cross-cluster edges) and robustness to exposure-mapping choices.

Finally, explain how you will detect strategic adversarial adaptation (e.g., attempts to circumvent the label) during the test.

Loading comments...

Browse More Questions

More Analytics & Experimentation•More Meta•More Data Scientist•Meta Data Scientist•Meta Analytics & Experimentation•Data Scientist Analytics & Experimentation