Rocchio Algorithm

Rocchio Algorithm

The Rocchio Algorithm is a classic technique for implementing Relevance Feedback. It updates an initial query vector by moving it toward the centroid of known relevant documents and away from the centroid of known non-relevant documents.

Rocchio Query Update

${\vec{q}}_m=\alpha {\vec{q}}_0+\beta \frac{1}{|D_r|}{\sum }_{\vec{d}\in D_r}\vec{d}-\gamma \frac{1}{|D_{nr}|}{\sum }_{\vec{d}\in D_{nr}}\vec{d}$

where:

• ${\vec{q}}_m$ — the modified query vector
• ${\vec{q}}_0$ — the original query vector
• $D_r$ — set of known relevant documents
• $D_{nr}$ — set of known non-relevant documents
• $\alpha ,\beta ,\gamma$ — weights (hyperparameters) controlling the balance between original intent, positive feedback, and negative feedback.

Vector Space Navigation

Imagine the Vector Space Model where points represent documents. A user’s query is also a point. If the user marks some results as “good,” Rocchio literally “nudges” the query point closer to those good results. Usually, we weigh relevant documents more heavily than non-relevant ones ($\beta >\gamma$).

Pseudo-Relevance Feedback (PRF)

Since users rarely provide explicit feedback, we often use Pseudo-Relevance Feedback:

1. Run initial search.
2. Assume the top $K$ documents are relevant (no $D_{nr}$).
3. Apply Rocchio to expand the query.
4. Run second search with ${\vec{q}}_m$.

Connections

• Context: Used within the Vector Space Model (VSM).
• Related concepts: BM25 (which typically doesn’t use vectorcentroids), Query Expansion.
• Modern link: Modern PRF methods often use Transformers to expand the query instead of vector arithmetic.

Appears In

• IR-L03 - Retrieval Models