Research On Personalized Recommendation System Based On ElasticSearch

In recent years,the development of science and technology has brought the continuous growth of information.The emergence of the recommendation system is to solve the problem of finding useful information for users difficult.For the individual needs of different users,recommendation system by analyzing users' preferences and historical records,finds the point of concern to the user,recommends to the user their most interested information,so as to meet the individual needs of the users.The problem of the increase of information is that it can not retrieve relevant information quickly.Therefore,based on the rapid retrieval nature of search engine,a recommendation system architecture based on ElasticSearch is proposed combining search engine with recommendation system.First of all,a brief introduction to the universal recommendation system is made,and the advantages and disadvantages of different recommendation algorithms are described.Secondly,this paper introduces the full-text retrieval ability of ElasticSearch and proposes a TFIDF keyword extraction algorithm based on time series.According to the historical record of different users,the user's keyword extraction is performed according to different weights,so as to solve the problem that search engines can not to meet the individual needs of the problem.Again,the machine learning algorithm is used to reorder the recall result of ElasticSearch,and three different machine learning algorithms are analyzed in depth.A user-item feature fusion method is proposed to extract positive and negative samples,and linear regression is Supported Vector Machine,as well as nonlinear model gradient boost tree.Then the ElasticSearch's recalling strategy and machine learning sorting strategy are combined to complete the entire system.The time series-based TFIDF keyword extraction is combined with the search capability of ElasticSearch to obtain the recalling result.Three kinds of machine learning algorithms are implemented,and the recalling results are reordered and returned to the user recommendation list.Finally,by crawling a company's customer data,the whole system is realized,and the class diagrams of the main parts and the sequence diagram of the whole system are concretely displayed.The AB-Test online comparison is performed on three different recommended algorithms.It is proved that the way of recalling the TFIDF keyword based on the time series has a relatively high click-through rate.