Reasearch On Personalized Recommendation Algorithm Of Agricultural Information Based On Group Users' Portrait

During the process of agricultural informatization in our country,a number of agricultural information is produced,thereupon the phenomenon of “information overload”arise,which brings more difficulties for the workers in agriculture to search for interested information.They are not able to get the satisfied information resources validly,timely and accurately.So it is an important approach for solving the “information overload” to excavate the Web log information produced by the interaction between users and system as well as grade information deeply and provide the recommending service of personalized information by predicting the users' interests and preferences through establishing the users' portrait.The agricultural users' portrait investigated in this paper mainly consist of sub-portrait of basic information,sub-portrait of content preference,sub-portrait of conversation and sub-portrait of grade.First of all,the sub-portrait of basic information and content preference are made use of to cluster then get the group users' portrait.Secondly,the technology of personalized recommendation is studied in the group by the separate application of sub-portrait of conversation and sub-portrait of grade.Lastly,the final recommendation is pushed to the user with weighting conformity.The main completed jobs are demonstrated as follows:(1)Research on agricultural group users' portrait.In allusion to the conditions of the less representativeness of interest model and difficulties in updating and preserving in the traditional personalized service system of agricultural information,the domain ontology of agricultural information field is built up based on the classification table on needs.Then the users' label data is achieved by data mining and the establishment of user portrait by combining the corresponding methods of mapping.Finally the storage,query and updating of user portrait are realized.Besides,the FCM algorithm based on the improvement of AP algorithm is applied to cluster the user portraits.Compared with traditional K-Means algorithm,improved K-Means algorithm and traditional FCM algorithm,the accuracy rate of predicting increases separately by 19.87%,9.75% and 7.25% while the accuracy rate of recommending increases separately by 11.48%,11.23% and 5.77%.(2)Research on recommendation technology based on the group users' sub-portrait ofconversation.In allusion to the problem of cold boot in the traditional collaborative filtering algorithm,the data mining is applied to recognize the information of users' conversation on the basis of streaming data produced from the interaction between user and system.According to the interest's characteristic of time quantum,the session set is formed by the partition of conversation combined with time,among which the similar conversation to the current activity conversation is found.Then the user in the traditional collaborative filtering algorithm is replaced by conversation.The system will analyze and predict the information the user in the current conversation pays most attention to and push them to the user as a recommendation.It is proved by experiment that this algorithm reaches to the best with an accuracy rate of 41.85%,a recall rate of 16.43%,a coverage rate of 25.28% and a popularity as 7.1746 when the neighbor is 120.Besides,this algorithm's function of HR@10 is improved by about 34.93% compared to traditional user-based collaborative filtering when recommending for the Top10.(3)Research on recommendation technology based on the group users' sub-portrait of ratings.This is aimed at the problem of low accuracy in the condition of extreme sparse data with the similarity computing method in the traditional collaborative filtering algorithm.To begin with,the information correlation coefficient is elicited using the related properties of discrete magnitude,based on which similarity values are calculated among users.Then the results of similarity are amended according to the close degree of users' interest and finally a comparative thorough computation of similarity is achieved.The experimental results show that the MAE in this method is 0.819002 when the data sparseness reaches to 0.9901,decreasing separately by 16.05%,14.35%,15.24% and 5.95% compared with COS-CF,ACOS-CF,PCC-CF and US-CF.It gives the proof that this algorithm is able to fit the environment with extreme sparse data.(4)Research on personalized recommendation model.In allusion to the condition of low information utilization in the traditional agricultural information service system,the design of agricultural information personalized recommendation system is based on the methods and technology studied comprehensively in this paper,which realizes some functions,such as agricultural information retrieval,personalized recommendation and so on.