Bayesian relevance feedback

Creator

Created

2025 Mar 10 15:28

Editor

Edited

2025 Mar 10 15:31

Refs

Optimum query Seeks to find a query that maximizes similarity to elevant documents while minimizing similarity with non-relevant documents

Let

C_r

denote set of relevant docs, Let

C_{nr}

denote set of non-relevant docs.

q_{opt} = \argmax_p [sim(q, C_r) - sim(q, C_{rn})]

q_{opt} = \frac{\sum_{d\in C_r} d_i}{|C_r|} - \frac{\sum_{d\in C_{rn}} d_i}{|C_{rn}|}

Compute the weight for term i using the log-odds ratio:

w_i = \log\left(\frac{p_i\,(1-q_i)}{q_i\,(1-p_i)}\right)

The updated query vector is then given by:

\vec{q}_{BRF} = \vec{q_0} + [w_1, w_2, \dots, w_n]