Multi-objective optimisation for information access tasks

M.J. Fisher, J.E. Fieldsend and R.M.Everson

Abstract

Information access tasks involve setting a large number of model parameters, for example choosing which terms from a document collection and which term weighting methods to use. Parameter values also have to be set, for example term weighting method parameters and thresholds. In addition, information access tasks frequently attempt to optimise multiple competing objectives, such as precision and recall. Here we introduce a multi-objective evolutionary algorithm framework for evaluating, analysing and improving performance in information access tasks by finding the sets of model parameters that simultaneously optimise both precision and recall. We present information access experiments using the TREC 7 and 8 collections for ad hoc query problems and the Cora collection for document classification. In particular, we compare the following term weighting functions: tf, idf, tf × idf, tf × idf/ndl, BM25 with standard parameter values, BM25 with parameters optimised for the document collection, and a novel neural network term weighting method. We find that the BM25 method with optimised parameter values outperforms all other methods including BM25 with standard values. The neural network method with optimised weights and biases also performs well. We also optimise the range of terms used for document classification and show how, for a particular information access task, to locate the optimum range of terms.

Gzipped postscript  (683 kb)     PDF  (5283 kb)