Ranking Search Results Based on User Preferences in the Absence of Personalized Statistics

Abstract

The article considers the algorithm of search results ranking according to the compliance with user's expectations due to the feedback on the data of search results pre-release, which allowed to reduce the time of search for the necessary information. A numerical experiment aimed at evaluating the effectiveness of the proposed approach on the example of real estate objects is presented. Aggregated data of real estate listings were used as data, and the results of user surveys were used to evaluate the relevance. TOPSIS and PROMETHEE methods were used as pre-ranking algorithms. The ranking results were combined considering their order in the output of both methods. To implement the next step of the algorithm, the pre-release list was partitioned into classes for which users selected a few objects they liked. Machine learning models were trained on the partitioned data. Based on their average accuracy and variance estimates, a naive Bayesian classifier model was selected and used for subsequent computations. The results of further experiments showed the possibility of taking into account personal preferences when organizing search and selection of objects of interest in the absence of personalized statistics on the example of real estate objects. As a result of the experiments, the time of searching a group of results of interest was reduced by 74 % on average.