Explain Harassment Surge
Problem Description
Suppose your monthly moderation analysis indicates a sharp increase in Harassment during the most recent month. A post is classified as violating if its probability_violating score exceeds 0.5. The monthly distribution of harassment is based on distinct posts viewed within each month.
As a Data Scientist specializing in content integrity, you are tasked with addressing this surge.
Questions
- Plausible Explanations: What are the most plausible explanations for this apparent surge in
Harassment? Consider both genuine changes in harmful user behavior and potential measurement artifacts. - Investigation Strategy: How would you investigate whether this observed increase is a true rise in harassment or an an artifact? Specifically, consider potential confounding factors such as changes in traffic mix, recommendation algorithms, content policies, underlying moderation model retraining, classification threshold adjustments, or data quality issues.
- Metrics and Statistical Checks: What specific metrics, segmentations, and statistical checks would you employ during your investigation? Your answer must explicitly address the impact of denominator effects, Simpson's paradox, selection bias, label drift, and model calibration.
- Interventions: If the increase is determined to be real, what product, policy, operational, or modeling interventions would you propose to mitigate the issue?
- Mitigation Experiment Design: If you were to test a proposed mitigation, how would you design an experiment or a quasi-experiment? What primary metrics and guardrail metrics would you define for this evaluation?
AI interview prep
Core dilemma
Distinguishing between a genuine increase in harmful user behavior and an artifact of the measurement system or platform changes, requiring systematic root cause analysis in a complex, dynamic environment.
3 tradeoffs and 3 stress tests locked
Interview prep
Unlock the interviewer X-Ray rubric, tradeoff matrix, stress tests, clarifying questions, and key concepts.
Explain Harassment Surge
Problem Description
Suppose your monthly moderation analysis indicates a sharp increase in Harassment during the most recent month. A post is classified as violating if its probability_violating score exceeds 0.5. The monthly distribution of harassment is based on distinct posts viewed within each month.
As a Data Scientist specializing in content integrity, you are tasked with addressing this surge.
Questions
- Plausible Explanations: What are the most plausible explanations for this apparent surge in
Harassment? Consider both genuine changes in harmful user behavior and potential measurement artifacts. - Investigation Strategy: How would you investigate whether this observed increase is a true rise in harassment or an an artifact? Specifically, consider potential confounding factors such as changes in traffic mix, recommendation algorithms, content policies, underlying moderation model retraining, classification threshold adjustments, or data quality issues.
- Metrics and Statistical Checks: What specific metrics, segmentations, and statistical checks would you employ during your investigation? Your answer must explicitly address the impact of denominator effects, Simpson's paradox, selection bias, label drift, and model calibration.
- Interventions: If the increase is determined to be real, what product, policy, operational, or modeling interventions would you propose to mitigate the issue?
- Mitigation Experiment Design: If you were to test a proposed mitigation, how would you design an experiment or a quasi-experiment? What primary metrics and guardrail metrics would you define for this evaluation?