Abstract
It is very hard, if not impossible to use analytical objective functions for optimization of personalized search due to the difficulties in mathematically describing qualitative problems. To solve such optimization problems, interactive evolutionary algorithms, which can make use of human preferences, are highly desirable. However, due to the lack of effective encoding methods, interactive evolutionary algorithms have been limited to numerically encoded optimization problems. In practice, however, linguistic terms (words) are the most natural expression of human preferences, and they are also commonly used to describe items in personalized search or E-commerce; therefore, language models better suit encoding, and the optimization of personalized search is converted into a dynamic document matching problem. To optimize word-described personalized search, we propose a novel interactive estimation of distribution algorithm. This algorithm combines a language model-based encoding approach, a Dirichlet-Multinomial compound distribution-based preference expression, and a Bayesian inference mechanism. The proposed algorithm is applied to two personalized search cases to demonstrate the capability of the algorithm in ensuring a more efficient and accurate search with less user fatigue.
[Display omitted]
•The presented IEDA employs a language model to encode candidate searched items.•The language model introduces social intelligence and reduces information loss.•The IEDA adopts Dirichlet-Multinomial distribution as its probabilistic model.•The probabilistic model is updated with Bayesian learning to track variable.•Then, a faster personalized search can be expected.