Information Retrieval · Cluster
Ranking
Scoring models and feature design.
- 01 BM25 Term Saturation and Document Length Normalization How the BM25 scoring function bounds the reward for repeated query terms, what the k1 and b parameters actually control, and where the model breaks down.
- 02 Disjunction-Max Combination for Multi-Field Scoring Why scoring each field separately and summing produces runaway scores, and how max-plus-tiebreaker combination keeps field evidence bounded.
- 03 Score Normalization and Rank Fusion Across Retrievers Why scores from different retrievers cannot be compared, how score distributions are normalized, and what reciprocal rank fusion gives up to avoid the problem.
- 04 Static Rank Priors and Their Effect on Ranking How document-level signals that ignore the query accumulate into a prior, how that prior is combined with query-dependent scores, and why it can dominate a weak match.
- 05 Query-Document Feature Extraction for Learned Ranking How match statistics, proximity windows, and coverage ratios are computed from position lists, and what each feature measures that a bag-of-words score cannot.
- 06 Click Features and Position Bias in Ranked Lists Why a click at rank one means something different from a click at rank nine, how examination models correct for it, and what inverse propensity weighting costs.
- 07 Pointwise, Pairwise, and Listwise Training Objectives What each learning-to-rank formulation optimizes, why pairwise ranking losses match the evaluation better than regression, and what listwise gradients buy.
- 08 LambdaMART and Metric-Driven Gradient Boosting How lambda gradients weight document pairs by their effect on NDCG, why the resulting tree ensemble is trained on pseudo-residuals, and where the heuristic gradient fails.
- 09 Score Calibration for Blended Ranking Why a ranker optimized for ordering produces scores with no absolute meaning, how to fit a monotone calibration, and what breaks when the query mix shifts.
- 10 Diversity-Aware Ranking and Subtopics How ambiguous queries are covered by selecting results that span subtopics, what maximal marginal relevance maximizes, and why diversity costs relevance.