1104335 - Implement Bayesian probability assignment in weighted rules scenario

Status: RESOLVED INCOMPLETE

(Content Services Graveyard :: Classification Engine, defect)

Description

[maxim zhilyaev]

Rules generated from moreover corpus have the form of:

rule: [cat1: probability, cat2: probability, ....]

The current algorithm implemented in ruleClassify chooses highest weighed category which is incorrect.

The correct formulation of probability for a given cat C is below:

P(C| R1 & R2 & R3) = P(R1|C) * P(R2|C) * P(R3|C) * P(C) / (P(R1) * P(R2) * P(R3))

Then all C above particular threshold are chosen, OR the most probable category is chosen.  Either algorithm is worth testting

Comment 1

[maxim zhilyaev]

This procedure does not seem to be necessary as simple selection of rules with precision above 85% seems to provide good overall precision recall in folding scenario.


Suggest putting it on back burner