Bot-Detection System Design for Comment Activity

Context

You are designing and evaluating a machine learning system to detect automated (bot) comment activity on a large-scale social platform. Bots are rare (e.g., <0.5% of comments) and adversarial. Your solution should balance safety (blocking bots) and user experience (minimizing false positives on humans), and it must work both offline and online.

Tasks

1. Labeling strategy with minimal ground truth

• Propose weak-supervision heuristics.
• Define a manual review sampling plan.
• Explain how to de-bias labels given extreme class imbalance (<0.5% bots).

2. Features across time scales

• Specify features: session burstiness, inter-comment intervals, entropy/diversity of targets, language signals, graph-based reciprocity, account/device/network signals.
• Indicate which features must be real-time vs. batch.

3. Model choice and calibration

• Compare linear models, tree ensembles, and sequence models.
• Describe how to calibrate posteriors (Platt scaling, Isotonic) and how to monitor calibration drift.

4. Thresholding by cost

• Define costs for FP (blocking a human) vs FN (missing a bot).
• Choose an operating point using precision–recall curves.
• Compute expected blocked-human-minutes at a chosen threshold given example rates.

5. Adversarial robustness

• Identify features least gameable, propose canaries, and drift detection.

6. Online safety net

• Outline shadow mode, backfill re-scoring, and human review queues.

7. Evaluation

• Offline: PR-AUC, recall at high precision, slice analysis.
• Online guardrails: abuse reports, creator retention, comment latency.

8. If FP becomes high

• Trace root cause with error analysis and propose a rollback/ramp strategy.