Learning to rank for information retrieval using layered multi-population genetic programming

To determine which documents are relevant and which are not to the user query is one central problem broadly studied in the field of information retrieval (IR). Learning to rank for information retrieval (LR4IR), which leverages supervised learning-based methods to address the problem, aims to produce a ranking model automatically for defining a proper sequential order of related documents according to the given query. The ranking model is employed to determine the relationship degree between one document and the user query, based on which a ranking of query-related documents could be produced. In this paper we proposed an improved RankGP algorithm using multi-layered multi-population genetic programming to obtain a ranking function, trained from collections of IR results with relevance judgments. In essence, the generated ranking function is consisted of a set of IR evidences (or features) and particular predefined GP operators. The proposed method is capable of generating complex functions through evolving small populations. LETOR 4.0 was used to evaluate the effectiveness of the proposed method and the results showed that the method is competitive with RankSVM and AdaRank.