Inferring probability of relevance using the method of logistic regression

摘要: This research evaluates a model for probabilistic text and document retrieval; the utilizes technique of logistic regression to obtain equations which rank documents by probability relevance as function query properties. Since infers from statistical clues present in texts queries, we call it inference. By transforming distribution each clue into its standardized (one with mean μ = 0 standard deviation σ 1), method allows one apply coefficients derived training collection other collections, little loss predictive power. The is applied three well-known information retrieval test results are compared directly particular vector space uses term-frequency/inverse-document-frequency (tfidf) weighting cosine similarity measure. In comparison, inference performs significantly better than (in two collections) or equally well third collection) tfidf/cosine model. differences performances models were subjected tests see if statistically significant could have occurred chance.