Research On Interactive Estimation Of Distribution Algorithms For Personalized Search With User Generated Contents

Efficient and accurate personalized search service can bring great convenience in the production and life.With the "information overload" of User Generated Contents(UGCs),such as: interactive behaviors,ratings,items' category tags,text comments,social network relationships,images or video information,etc.,it has become increasingly complex.Personalized search with UGCs is currently a research hotspot in the field of big data analysis and personalized services.This problem is difficult to establish a clearly defined mathematical model and objective function,and its analysis and evaluation decision-making process is subjective,inconsistent and fuzzy.It is a complex optimization problem with qualitative indicators.Interactive Evolutionary Computations(IECs)that integrate users' interactions and intelligent evolutionary optimization algorithms are effective approaches to solve the optimization problems with qualitative indicators.However,it is a serious challenge to design efficient interactive evolutionary optimization strategies integrateing multi-source,multi-mode,heterogeneous,and unstructured data in IECs for effectively solving the personalized search tasks with UGCs.To this end,this paper studies interactive Estimation of Distribution Algorithms(IEDA)for personalized search with UGCs.The main contents include the following four points:(1)RBM preference model-assisted IEDA based on user interactions.Personalized search with UGCs is a typical qualitative index discrete variable optimization problem.When using intelligent optimization algorithms,user preference and evaluation surrogate models should be firstly designed.Therefore,combing the powerful feature extraction capability of Restricted Boltzmann Machine(RBM)with the optimization capability of Estimation of Distribution Algorithms(IEDA),an EDA based on RBM-assisted surrogate and probability models is presented.This algorithm is extended to design an RBM preference surrogate model that considers users' historical interaction behaviors and items' category tags,that is,the training samples with users' preference information are formed according to users' interaction behaviors to screen item sets with long interaction time or high user evaluations.Binary encoding of items' category is regarded as the input of the RBM preference cognitive model to extract users' preference features based on the training samples.The EDA probability model for the preference feature distribution and RBM energy function-based fitness estimation surrogate model are constructed according to the RBM preference model.An IEDA integrating RBM preference surrogate model is constructed and its complexity is analyzed.The application of the algorithm in complex discrete functions and Movie Lens personalized search proves the effectiveness of EDA integrating RBM surrogate model and IEDA with user interactions-driven RBM preference surrogate model.(2)Enhanced dual RBM-driven IEDA with users' implicit preference in UGCs.Although Research content(1)used users' interaction behaviors,ratings and items' category information to improve IEDA,it does not make full use of users' social network information and the implicit positive and negative preference features in users' evaluations.In view of this,based on the research content(1),an IEDA with the dual-RBM surrogate model based on the positive and negative preferences is further studied.According to user interaction behaviors in UGCs,such as interaction duration and ratings,explicit preference and implicit preference information are obtained to screen dominant and inferior groups to construct a dual-RBM user preference model that simultaneously recognizes positive and negative preferences to extract user preference features more precisely.By using the marginal probability distribution of the features of the positive RBM preference model,the EDA probability model is presented.The fitness estimation model is constructed based on the social network information and the energy function of the weighted positive and negative RBM preference models,and an efficient IEDA algorithm is designed.A large number of experiments in the Amazon datasets show that the proposed algorithm can not only effectively enhance the performance of personalized search,but also reduce the burden of user evaluations to improve users' interactive search experience.(3)Enhanced IEDA with multi-source heterogeneous UGCs-based RBM surrogate.The above researches only consider users' interaction behaviors,items' category tags and social network information in UGCs,and do not consider the large amount of users' text comments in UGCs that contains a large number of userss' implicit preferences.In view of this,based on the research content(2),an RBM preference surrogate IEDA algorithm for multi-source heterogeneous UGCs data is proposed.Considering users' ratings,category tag,text comments and social relationships in UGCs,their mathematical descriptions are given.The vectorized representation of multi-source heterogeneous text data is obtained by using a doc2 vec model.An RBM parallel double-input user preference model with both discrete category features and continuous semantic features is constructed based on searched objects' category tags and users' evaluation text vectors.A surrogate model based on the RBM user preference and multi-similar user social relations is designed to estimate the individual fitness of feasible solutions.The parameters of the RBM preference model,probability model and fitness estimation function are dynamically updated to realize an IEDA that accurately tracks the user preferences to improve the efficiency of personalized search.The application results of the algorithm in the Amazon datasets show that the proposed algorithm can better predict users' preferences,dynamically track the changes of users' interests,effectively reduce users' evaluation burden and improve the accuracy and efficiency of personalized search.(4)Enhanced RBM-driven IEDA with attention mechanism and multi-source heterogeneous UGCs.After integrating multi-source heterogeneous data in Research content(3),the decision variables include both items' category features and text's implicit features.Obviously,these features have different effects on users' preferences and different contributions on the RBM-based preference surrogate model.For this reason,the RBM preference surrogate-assisted IEDA with the attention mechanism is further studied to extract the importance of decision variables.According to items' categories,text comments and collaborative information in UGCs,a doc2 vec model and multi-hot encoding mechanism are used to integate multi-source heterogeneous data to design the RBM-based attention weight module.Combining the attention weights of users' preference features,an RBM user preference model is construct to describe search objects from multiple angles.The EDA distribution estimation probabilistic model based on the attention weights of preference features and the RBM-based user preference surroagte model are designed in the IEDA framework.According to new user interactions and UGCs data,the model management mechanism is used to update the RBM-based user preference model that integrates multi-source heterogeneous data and AM to dynamically track users' preferences.The application of the algorithm in the Amazon datasets shows that the proposed IEDA further improves the ability to fit users' preferences and the accuracy of personalized search.The above research contents adopt intelligent optimization algorithms for the personalized search with UGCs to construct multiple RBM cognitive models reflecting users' preferences by using UGCs from different aspects.Based on the probability distribution and energy function of RBM model on preference features,the construction mechanism of the EDA probability model and fitness surrogate model based on the RBM preference model are given in different scenarios.Furthermore,an efficient IEDA is designed to solve the personalized search problem with UGCs.The application of complex functions and actual Amazon personalized search proves the effectiveness of the proposed algorithm.In this paper,there are 31 figures,22 tables and 221 references.