Investigate Harassment Surge and Mitigation

Question

You are on a content integrity team. A monthly moderation review shows a sharp increase in the Harassment violation type over the latest month. Assume a post is flagged as violating when probability_violating > 0.5, and the monthly distribution counts distinct posts viewed during each month. A single post may fall under multiple violation types. As a Data Scientist, work through the following:

1. Frame the surge precisely. Separate a volume increase (a larger count of harassment posts) from a rate increase (a higher share or prevalence). Which prevalence definitions and denominators would you compare, and across which time windows?
2. Plausible explanations. Enumerate the most likely causes, covering both real-world and measurement-related explanations, including:
   • an actual rise in abusive behavior
   • shifts in traffic mix across surfaces, regions, languages, or creator cohorts
   • seasonality or outside events
   • organized attacks or repeat offenders
   • changes to policy definitions
   • changes in model thresholds or model versions
   • calibration drift in the classifier
   • data-quality, logging, or backfill problems
3. Investigate real vs. artifact. Outline a concrete investigation plan. Specify the prevalence metrics, segments, and additional datasets you would request (e.g. human-review labels, user reports, enforcement logs, model-version metadata). Your statistical reasoning must explicitly cover denominator effects, Simpson's paradox, selection bias, label drift, and model calibration, and explain how you would check model-driven metrics against human-reviewed labels.
4. Propose interventions (if the surge is real). Recommend product, ranking, policy, operational, and ML remedies, separating short-term containment from longer-term fixes.
5. Evaluate the mitigation. Design an experiment or quasi-experiment to test one chosen intervention. Specify the primary metrics, guardrail metrics, unit of randomization (or rollout design), and the key tradeoffs involving false positives, fairness, and user experience. Overview: A Data Scientist onsite analytics question at Airwallex: a monthly moderation report shows the Harassment violation type spiking, and you need to determine whether the surge is real or a measurement artifact, then propose and test mitigations. It tests diagnostic analytics, prevalence-metric design, denominator effects, Simpson's paradox, selection bias, label drift, model calibration, causal inference (ITS/DiD), and experiment design with guardrails.