A relevance feedback method in vector‑space retrieval that updates an original query vector by forming a weighted linear combination of the original query and the centroids (average vectors) of judged relevant and nonrelevant documents, typically q' = αq + β(1/|R|)Σv_r − γ(1/|NR|)Σv_nr with tunable weights α, β, γ.

Principle

Move the query vector toward the centroid of judged relevant documents and away from nonrelevant ones in the vector space; linear combination weights determine the extent of adaptation and trade off original intent with feedback evidence.

Demonstration

Illustrative scenario → User submits a query and marks two returned documents as relevant and one as nonrelevant. Recognition → Compute centroid of the two relevant document vectors and centroid of the nonrelevant vector. Action → Update the query vector using chosen α, β, γ and re-run ranking. Consequence → Documents near the relevant centroid increase in score; if feedback is accurate, precision for similar documents rises.

Misapplication

Applying Rocchio with very small or noisy feedback sets (e.g., a single mistakenly labeled relevant document) and large β leads to query drift. The error appears plausible because the centroid seems informative; the semantic error is treating sparse feedback as representative of the user's full information need.

Consequence

When user judgments are reliable and vector representations capture topical structure, Rocchio can focus retrieval and improve effectiveness; when feedback is sparse or noisy it can degrade performance by amplifying idiosyncratic signals.

Reversal

If the vector space poorly represents document semantics (e.g., synonymy unresolved) or feedback reflects diverse intents, linear centroid shifting can be ineffective or harmful. In such cases nonlinear or learning‑to‑rank approaches that use richer features may be preferable.

Boundary

Clearly within: term‑weighted vector‑space models applying linear centroid-based query updates. Boundary case: pseudo‑relevance feedback uses top‑k results as R without explicit user labels. Clearly outside: feedback mechanisms that train discriminative rankers or use neural representation fine‑tuning rather than linear vector updates.

Semantic Tension

Tension between rapid, interpretable linear query adjustment (efficiency, transparency) and more flexible but complex supervised reranking methods that can model interactions and nonlinearities in feedback.

Synthesis

Rocchio formalizes a geometric intuition—pulling the query toward relevant examples and pushing away from negatives—offering simple, transparent feedback adaptation whose success depends on the representational quality and reliability of the feedback.