Fusion Normalization

Standardizes scores by measuring how many standard deviations a value lies from the mean, centering the distribution around zero.

• Formula: score_norm = (score - μ) / σ
• Strength: Robust to outliers compared to Min-Max; handles varying score magnitudes well.
• Weakness: Assumes a roughly Gaussian distribution; can produce negative scores that require downstream handling.
• Best for: Merging results from systems with unknown or highly variable score ranges, such as combining BM25 with cosine similarity.

Divides each score by the sum of all scores in the result set, converting them into a probability distribution where all values sum to 1.

• Formula: score_norm = score / Σ(scores)
• Strength: Produces an easily interpretable proportional weighting.
• Weakness: The normalized value of a document depends on the scores of other documents in the set, making scores non-portable across queries.
• Best for: Weighted Sum Fusion where you need a convex combination of normalized scores.

Discards raw scores entirely and assigns a new score based solely on a document's position in the ranked list. This is the conceptual basis for Reciprocal Rank Fusion (RRF).

• Approach: Assign score = 1 / (k + rank) where k is a constant (often 60).
• Strength: Completely immune to score calibration issues; treats all retrieval sources equally.
• Weakness: Discards the magnitude of relevance signals; a marginally relevant top result is treated identically to a highly relevant one.
• Best for: Heterogeneous retrieval ensembles where score magnitudes are non-comparable or untrustworthy.

Models the empirical score distribution from each retriever and uses the fitted cumulative distribution function (CDF) to map scores to quantiles or probabilities.

• Approach: Fit a Gaussian, log-normal, or beta distribution to historical score data, then use the CDF for normalization.
• Strength: Produces well-calibrated, probabilistic scores that are directly comparable across systems.
• Weakness: Requires sufficient historical data to fit a reliable distribution; adds complexity to the pipeline.
• Best for: High-stakes production systems where Score Calibration is a formal requirement and historical query logs are available.

A pragmatic pre-processing step that truncates each result list to a fixed depth before normalization, discarding low-scoring tail results that add noise.

• Approach: Keep only the top 100 or 1000 documents per retriever before applying Min-Max or Z-Score.
• Strength: Reduces computational overhead and prevents long-tail noise from distorting normalization parameters like min, max, or mean.
• Weakness: Risk of discarding relevant documents if the clipping threshold is too aggressive.
• Best for: Multi-Stage Retrieval pipelines where the fusion step feeds into a more precise Cross-Encoder Reranking stage.

A linear scaling technique that transforms raw scores into a fixed range, typically [0, 1]. The formula (score - min) / (max - min) preserves the original distribution's shape while making scores from different retrievers directly comparable.

• Sensitivity: Highly susceptible to outlier scores, which can compress the rest of the distribution.
• Requirement: Needs a known or estimated minimum and maximum score per query.
• Application: Common in pre-fusion pipelines where score ranges are predictable.

A statistical normalization method that rescales scores based on their mean and standard deviation: (score - μ) / σ. This centers the distribution around 0 and expresses each score in terms of standard deviations from the mean.

• Robustness: Less sensitive to outliers than Min-Max, but assumes a roughly normal distribution.
• Output Range: Unbounded; scores can be negative.
• Best For: Retrieval systems where score variance is a meaningful signal of confidence.

A linear combination method where each retriever's normalized score is multiplied by a pre-defined weight reflecting its importance. The final score is the sum of these weighted contributions.

• Formula: FinalScore = w1*score1 + w2*score2 + ...
• Dependency: Requires prior score normalization for meaningful weighting.
• Tuning: Weights can be set heuristically or learned via a Learning to Rank (LTR) model on validation data.

The process of transforming raw model outputs into well-calibrated probabilities that reflect the empirical likelihood of relevance. Unlike simple normalization, calibration aims for statistical correctness.

• Platt Scaling: Fits a logistic regression model on raw scores.
• Isotonic Regression: A non-parametric method that learns a monotonic mapping.
• Importance: Enables meaningful score comparison and fusion when retrievers have different confidence characteristics.

A strategy that combines results from multiple heterogeneous retrieval systems—dense vectors, sparse keywords, and metadata filters. Fusion normalization is the critical glue that makes ensemble retrieval possible.

• Goal: Improve overall recall and robustness against individual system weaknesses.
• Challenge: Each retriever produces scores on a different, uncalibrated scale.
• Solution: Apply Min-Max, Z-Score, or RRF before merging into a single candidate pool for re-ranking.