weighted-engagement-scoring

Normalize each metric to the [0, 1] range first, then compute a weighted sum where the weights sum to 1.0. Each weight needs a documented rationale, and the result should be validated by manually inspecting the top and bottom ranked items.

Use log min-max normalization for power-law distributed counts like scores and comment volumes, since a single viral outlier would otherwise dominate the range. Use robust percentile-based min-max for moderate skew and standard min-max for bounded metrics like ratios.

Equal weights are acceptable only when explicitly justified with the rationale that no signal is a priori more important. Arbitrary or undocumented weights make the analysis irreproducible, so every weight choice needs a one-sentence justification.

This happens when metrics are combined without normalization, letting the metric with the largest numerical range control the result. It can also occur when two highly correlated metrics (r > 0.85) double-count one signal, so check the correlation matrix before weighting.

Avoid it when only one metric is available, when ranking users rather than content, or when comparing scores across different platforms without recalibrating weights. Also avoid presenting it as a quality metric without including a sentiment signal, since high engagement includes controversy and rage-bait.

Run a sensitivity analysis by perturbing each weight by +/-10% and measuring top-N overlap. Above 90% overlap means the ranking is robust, 50-70% means results depend heavily on weight choices, and below 50% means the composite score is unreliable.