You are a Data Scientist in a Trust & Safety team for a short-video platform (similar to TikTok/Reels).
The team asks: “How would you design risk metrics, and how would you decompose (拆解) them?”
Context
The platform faces multiple risk sources and controls:
-
Risk types
: policy-violating content (nudity/violence/hate), spam/scams, bot activity, account takeover, misinformation.
-
Detection & enforcement signals
: ML model predictions, proactive human review, reactive user reports, and enforcement actions (remove, downrank, age-gate, suspend).
Task
-
Propose a
metric framework
to measure “risk” at the platform level.
-
Define
primary metric(s)
and
diagnostic / guardrail metrics
.
-
Specify exact definitions (numerator/denominator), unit (per view/per user/per content), and time window (e.g., daily in UTC).
-
Explain what it means to
decompose
the risk metric (“拆指标”), and provide a concrete
metric tree / breakdown
that helps identify
why
risk increased or decreased.
-
Show at least two different decomposition approaches (e.g., by funnel stage vs. by segment).
-
List the key
data/measurement pitfalls
and how you would address them.
-
Examples: label delay, selection bias from user reports, changing enforcement policies, feedback loops from downranking/removal, duplicated content, and seasonality.
Your answer should be actionable for ongoing monitoring and root-cause analysis, not just high-level ideas.