Learning to rank based on modified genetic algorithm

With the growing amount of documents in the search index of information retrieval systems, the problem of ranking documents becomes crucial. The modern state of the problem leads to the point where machine learning becomes the most efficient way to optimize the ranking function. In this article investigated ranking function in information retrieval systems (IRS) and learning to rank problem. During the learning to rank process, IRS is defining the weight coefficients for simple rankers. The conducted researches are showing the approach for learning to rank problem LTR-MGA utilizing the hybrid method based on modified genetic algorithm and the Nelder-Mead method. This approach can be used to optimize a graded-metrics of ranking, such as NDCG. The efficiency of proposed method was proved, based on researches performed on LETOR data sets. The value of ranking quality measures was significantly increased after learning to rank process. Also the usage of modified genetic algorithms leads to reduction of time required for learning to rank comparing to traditional genetic algorithm.