Score Normalization

A linear rescaling technique that transforms raw scores into a fixed range, typically [0, 1], using the formula:

score_norm = (score - min) / (max - min)

• Strengths: Simple to implement, preserves the original distribution shape, and guarantees a bounded output range.
• Weaknesses: Highly sensitive to outliers; a single extreme score can compress the rest of the distribution into a narrow band.
• Best for: Retrieval pipelines where the score distributions are known to be stable and free of extreme outliers.

Transforms scores based on their statistical properties by centering the distribution around a mean of 0 with a standard deviation of 1:

score_norm = (score - μ) / σ

• Strengths: Robust to outliers compared to Min-Max, handles negative scores naturally, and does not assume a bounded range.
• Weaknesses: Assumes a roughly Gaussian distribution; can produce unbounded values that require clipping before fusion.
• Best for: Systems where score distributions are approximately normal and you need to compare relative standing rather than absolute magnitude.

Converts a set of raw scores into a probability distribution where all values sum to 1.0:

score_norm = score / Σ(scores)

• Strengths: Produces an intuitive probabilistic interpretation, making it easy to reason about relative importance across the candidate set.
• Weaknesses: The normalized value of a document depends on the scores of all other documents in the result set, making it query-dependent and non-comparable across queries.
• Best for: Late-stage re-ranking where you need a softmax-like distribution over a fixed candidate pool.

Discards raw score magnitudes entirely and assigns a new score based solely on a document's rank position within its result list:

score_norm = 1 / (rank + k)

• Strengths: Completely immune to score distribution differences between retrieval systems; the most robust method when raw scores are non-comparable.
• Weaknesses: Destroys the granularity of the original scores; documents with vastly different raw relevance but adjacent ranks receive nearly identical normalized scores.
• Best for: Reciprocal Rank Fusion (RRF) and scenarios where the retrieval subsystems produce scores on fundamentally different scales.

Maps raw scores to a target probability distribution, such as a Gaussian or log-normal distribution, using quantile mapping or histogram equalization:

• Strengths: Forces all subsystems to conform to an identical statistical shape, enabling mathematically sound linear combination.
• Weaknesses: Computationally more expensive; requires estimating the empirical CDF of each subsystem's scores, often from a calibration set.
• Best for: High-precision fusion pipelines where the cost of miscalibration is high, such as financial or medical search applications.

Uses a representative sample of historical queries to pre-compute normalization parameters (min, max, mean, standard deviation) offline, which are then applied at query time:

• Strengths: Avoids the latency of computing normalization statistics on the live result set; parameters are stable and auditable.
• Weaknesses: Sensitive to distribution drift; if the nature of queries or documents changes over time, the calibration becomes stale.
• Best for: Production systems with strict latency budgets where the retrieval score distributions are stationary.

A linear rescaling technique that maps raw scores into a fixed [0, 1] range. The formula subtracts the minimum observed score and divides by the range (max minus min).

• Formula: (score - min) / (max - min)
• Strength: Simple, interpretable, and preserves the original distribution shape
• Weakness: Highly sensitive to outliers; a single extreme score compresses all other values
• Use case: Best when score distributions are roughly uniform and outlier-free

This is often the default choice before applying Weighted Sum Fusion or CombSUM.

A statistical normalization that rescales scores based on their mean and standard deviation, centering the distribution around zero.

• Formula: (score - μ) / σ
• Strength: Robust to outliers compared to min-max; handles varying distribution widths
• Weakness: Assumes a roughly Gaussian distribution; negative scores can be unintuitive
• Use case: When score distributions from different retrievers have vastly different variances

Z-scores enable fair comparison when one retriever produces tightly clustered scores while another produces widely dispersed values.

A rank-based fusion algorithm that bypasses score normalization entirely by operating on rank positions rather than raw scores.

• Formula: Σ 1 / (k + rank_i) where k is a constant (typically 60)
• Strength: No calibration needed; inherently handles incomparable score scales
• Weakness: Discards magnitude information; a document ranked 1st by a weak signal gets the same weight as 1st by a strong signal
• Use case: When score distributions are unknown, non-linear, or fundamentally incomparable

RRF is the default fusion method in Elasticsearch Hybrid Search and many production systems precisely because it eliminates the normalization problem.

A score aggregation method that multiplies the sum of normalized scores by the number of retrievers that returned the document, boosting consensus.

• Formula: (Σ normalized_scores) × count_of_retrievers
• Strength: Rewards documents found by multiple independent systems, reducing false positives
• Weakness: Requires proper score normalization first; can over-boost mediocre documents found by all systems
• Use case: When you have three or more heterogeneous retrievers and want to emphasize agreement

CombMNZ builds on CombSUM by adding a multiplicative consensus factor, making it particularly effective when retrieval diversity is high.

A machine learning approach that trains a model to learn optimal combination weights from labeled relevance data, replacing hand-tuned normalization and fusion formulas.

• Architecture: Typically a small neural network or gradient-boosted trees taking raw scores as features
• Strength: Captures non-linear interactions between retrieval signals; adapts to domain-specific patterns
• Weakness: Requires labeled training data; can overfit to training queries; less interpretable
• Use case: High-volume production systems with click-through data or explicit relevance judgments

Learned fusion subsumes both normalization and combination into a single optimized function, often outperforming static formulas when sufficient training data exists.

A dynamic approach that adjusts fusion weights per query based on query intent classification, rather than using static global weights.

• Mechanism: Classify query as navigational, informational, or transactional; boost sparse or dense accordingly
• Navigational queries: Increase Lexical Boosting weight for exact term matching
• Informational queries: Increase Semantic Boosting weight for dense vector similarity
• Strength: Adapts to the specific retrieval needs of each query type

This technique requires score normalization to be applied before the dynamic weights are assigned, ensuring the boosted scores are comparable.