x-ranking-engine

The Heavy Ranker calculates feed ranking scores by applying isotonic calibration to Multi-Task Learning action probabilities, then multiplying them by assigned action weights to compute the final candidate score.

MaskNet and Phoenix architectures process user and tweet features through Multi-Task Learning heads, outputting raw probabilities for actions like Like, Reply, and Author Reply before isotonic calibration adjusts them.

You adjust the Author Reply weight in the ranking engine configuration, then observe the recalculated weighted score to see how the candidate's final ranking position changes relative to other candidates.

Isotonic calibration is a non-parametric technique needed to adjust raw Multi-Task Learning probabilities into accurate action likelihoods before applying weights for final ranking score computation.

Yes, you can use this to diagnose ranking behavior by tracing feature interactions through the Heavy Ranker and validating how weight adjustments impact timeline diversity and recency.

Candidate ranking scores change after probability calibration because isotonic calibration adjusts the raw Multi-Task Learning probabilities, which directly alters the weighted sum producing the final score.