Interactive Estimation Of Distribution Algorithm For Personalized Search

As such a kind of qualitative optimization problems that are of no explicitly defined objective functions,personalized search is difficult to solve but widely exists.Traditional evolutionary optimization algorithms that require explicitly defined objective functions are not applicable anymore,while interactive evolutionary optimizations by merging user interactions can be a suitable solution to these aforementioned problems.In most cases,the existing studies on interactive evolutionary optimization-assisted personalized search model searched targets as combinations of limited attribute variables.Most of them adopt numerical encoding schemes to turn the search into combinatorial optimization and then solve it with genetic algorithms(GAs).However,none of these studies has ever tried any other evolutionary optimization mechanisms or utilized any searched-target-related domain knowledge.Moreover,they ignore an enormous amount of textual information and the unique needs of privacy protection in personalized search.Motivated by these issues,we carry out corresponding research on developing more effective interactive evolutionary algorithms.The main contents are as follows:(1)Domain-Knowledge-Driven Fast Interactive Estimation of Distribution Algorithm: when dealing with numerical personalized search problems,the subspace is first obtained by splitting solution space according to the queries input by the active user.With the help of user search records and the searched items,the subspace is then reduced by Naive Bayesian probabilistic model into the preferred initial search space,i.e.a combination of user-preferred variables.By quantifying the preference which is reflected by user interactive behaviours with interval numbers,the preference surrogate is consequently constructed based on RBF Neural Network to estimate individuals' fitness and conduct selection operation.The Bayesian-probabilistic-model-assisted interactive estimation of distribution algorithm is presented and its superiorities in improving search efficiency and alleviating user fatigue are experimentally demonstrated.(2)Language-Model-Assisted Interactive Estimation of Distribution Algorithm(LM-IEDA): when solving textual personalized search problems,to preserve the rich semantic information in texts,the LM-IEDA first obtains the Doc2Vec-based vector expression of texts with the help of the research achievements of Natural Language Processing(NLP).Using the vocabulary and word frequencies,textual personalized search is then turned into dynamic text matching.The presented LM-IEDA adopts the Dirichlet-Multinomial compound distribution describing word frequencies as IEDA's probabilistic model and employs the Bayesian inference to update it.According to the application of the proposed algorithm in the book and movie search,its effectiveness is experimentally illustrated.(3)Collaborative/Content-Hybrid-Filtering-Based Dual-Probabilistic-Model Interactive Estimation of Distribution Algorithm(DPM-IEDA): based on content(2),the DPM-IEDA is constructed through integrating the collaborative filtering with the content-based techniques of personalized recommendation to fully utilize social search information.The user preference is first estimated according to many similar users' preference with collaborative filtering so that the preference-estimation-based probabilistic model can be obtained.Based on the user preference reflected in interactive behaviours,the IEDA's probabilistic model is consequently constructed to sample preferable items to generate the evaluation list.The aforementioned two probabilistic models are dynamically updated with user evaluated items.When applying the DPM-IEDA into movie and TV plays personalized search,the algorithm's high efficiency of helping users locate satisfactory items is proven.(4)Privacy-Preserving Enhanced Federated Learning with Asynchronous Mode and Temporally Weighted Aggregation: as the prospective study of the following privacy preserving interactive evolutionary computation,the research on enhanced federated learning(FL)is carried out for addressing the high communication cost problem in existing FL algorithms.Inspired by the characteristics of local deep neural networks adopted in FL,the asynchronous mode is first proposed.Then,the temporally weighted aggregation is developed to fully utilize history information for an effective aggregation.The enhanced FL algorithm is thus proposed by integrating these two improved strategies.Finally,two applications including a CNN-based MNIST task and an LSTM-based HAR task show that the proposed algorithm is of lower communication costs and higher learning accuracy.(5)Privacy-Preserving Federated-Learning-Dual-Probabilistic-Model Interactive Estimation of Distribution Algorithm(FL-DPM-IEDA): based on(3)and(4),the FLDPM-IEDA is constructed to fulfil privacy-preserving requirement in hybrid personalized search.The privacy-preserving Federated-SVD(Singular Value Decomposition)is first developed by integrating SVD-based collaborative filtering techniques into the FL framework introduced in(4).Based on Federated-SVD,social preference probabilistic model is then proposed to initialize search space and population.Emerging with DPM-IEDA introduced in(3),the privacy preserving FL-DPM-IEDA is finally achieved,and its application in privacy-preserving hybrid personalized search proves its efficient search in privacy protection settings.The advantages of all these proposed algorithms have been experimentally demonstrated in corresponding personalized search applications.The study here can not only enrich the existing interactive evolution optimization theory but also promote the application of the above theories and methods in the optimization of qualitative optimization such as personalized search problems.Thus,they directly serve the development of the national economy and society.