Font Size: a A A

Research On Multidimensional Intelligent Recommendation Based On Social Behaviors

Posted on:2023-02-27Degree:DoctorType:Dissertation
Country:ChinaCandidate:H T DingFull Text:PDF
GTID:1528306914958519Subject:Management Science and Engineering
Abstract/Summary:PDF Full Text Request
As an effective tool to solve information overload,recommendation system has become a basic system capability of Internet platform and integrated into every interaction of users.Social behavior is an important kind of implicit feedback data reflecting users’ interests and preferences in the field of recommendation system.It is of great value to alleviate the problem of sparse data in recommendation system and improve the performance of recommendation system.With the comprehensive popularization of UGC(user produced content),users can easily publish,like,comment and forward content on social networks,resulting in a large amount of social behavior data.Social behavior data includes interactive data of log statistics,as well as text,image and other content data,which contains rich information and knowledge.Therefore,how to mine the user preferences hidden in social behavior data,build a more accurate user portrait and improve the performance of recommendation system are the key and difficult problems to be solved by the Internet platform.Based on this,according to the different data modes of social behavior,from the perspectives of interactive data mining and content data mining,this paper studies the intelligent recommendation based on social behavior by using scientific methods such as reasoning modeling,machine learning,deep learning,multimodal fusion and Bayesian model,in order to explore the related problems of social behavior data mining and its application in recommendation system.The main research contents of this paper include the following four aspects:(1)Using grounded theory to build social behavior recognition model and explore the internal driving factors of social behavior;Analyze the relationship between different social behaviors,put forward the correlation hypothesis and test it;Quantitative analysis studies the correlation between different social behaviors and users’ interests and preferences.(2)From the perspective of interactive data mining of social behavior,this paper studies the implicit recommendation of mixed social behavior.Firstly,an interest prediction model is proposed,and the model parameters are quantified by using social behavior interaction data.Based on the interest model,this paper constructs a machine learning based implicit recommendation model by using the method of matrix decomposition theory,and makes an empirical test on the model.(3)From the perspective of social behavior content data mining,this paper studies the construction of user portrait model based on multimodal content data.Explore the feature extraction method of social behavior text and image data,establish a user portrait model with multi-modal joint representation through in-depth learning,realize the feature extraction of user multi-point of interest and fine-grained portrait,and evaluate the effect of the model.(4)Based on the dynamically updated user portrait,this paper constructs an intelligent dynamic recommendation system centered on user behavior,explores the method of capturing user interests in real time and generating candidate sets quickly,and obtains the optimal recommendation sequence by using Bayesian ranking learning network.Aiming at the problem of user interest migration,the user portrait is updated in real time based on time-varying convolution window to realize intelligent dynamic recommendation,and the performance of the recommendation system is evaluated experimentally.The main innovations of this paper are as follows:(1)It identifies the internal driving factors of social behavior,reveals the correlation between different social behaviors,and obtains the correlation degree between social behavior and user preferences.The degree of interest and preference hidden behind different social behaviors is different.Users’ interest in items is the basis of the recommendation model.Previous studies on interest prediction models are mostly based on a single social behavior,or simply weighting different behaviors,which is subjective and lack of data support.Therefore,the analysis and exploration of users’ social behavior is particularly important.Using the method of grounded theory,this paper identifies the internal driving factors of users’ social behavior;Based on the statistical analysis and hypothesis test of real social media data,the correlation between different social behaviors is revealed;Taking purchase behavior as dependent variable and various social behaviors as independent variables,the correlation between social behavior and user preference is obtained through multiple regression analysis.The results provide theoretical support for using a variety of social behavior data to predict user interest and improve the effect of recommendation model.(2)This paper analyzes the implementation mechanism of interest prediction of social behavior optimization,puts forward the interest model of mixed social behavior,and constructs the argot recommendation model based on machine learning.As an implicit feedback data reflecting user preferences,the mining and application of social behavior is of great value ’to the recommendation system.This paper combines social behavior with display feedback data.Based on utility theory,an interest model of mixed social behavior is proposed,and the parameters of interest model are quantified by using social behavior interaction data.Based on the interest model and the matrix decomposition theory,a machine learning based implicit recommendation model is constructed.By extracting the hidden factor vector of users and items,the dimension of the original calculation vector is reduced and the vector density is increased,so as to improve the efficiency of similarity calculation and the accuracy of recommendation results.In the process of recommendation matching,the complexity of recommendation matching is reduced by fuzzy weighted clustering,and the diversity and richness of recommendation effect are maintained by controlling the number of clusters and the number of items selected in each cluster.The experimental results on UBA user behavior data set show that the experimental performance of the proposed interest prediction model and implicit recommendation model under each index has obvious advantages over the comparison method.(3)This paper reveals the value realization of social behavior content data in user portrait,and constructs a multi-modal data fusion user portrait model based on the characteristics of social behavior content.User portrait is the user’s interest preferences obtained by analyzing user behavior,which is an important basis for recommending to users.At present,Internet enterprises generally use the linear model based on Feature Engineering to construct user portraits.There are some problems,such as single data source,limited representation ability,manual extraction of discrete features,static portraits and so on.To solve the above problems,the paper first explores the value of social behavior content data to user portrait,and extracts the interest features of text content combined with TF-IDF algorithm and word2vec shallow neural network;The DPM target detection method is used to extract the effective area of the image,the hidden user interest points in the image content are fully mined based on the multi-scale attention mechanism,and the improved resnet-50 convolution network is used to extract the image interest features.Secondly,the user portrait model of multi-modal data fusion is constructed by using deep learning.Aiming at the semantic consistency between text mode and image mode,a self supervised multimodal feature confrontation learning network is proposed to make the modes of text feature and image feature tend to be consistent;The first mock exam is implemented by HDBSCAN method,and the multi modal clustering is implemented to represent the interest vectors of the unified modal.The experimental results on the SMP user portrait technology evaluation standard data set show that the evaluation index of the proposed multi-modal fusion method surpasses other multi-modal fusion and single-modal comparison methods.(4)The implementation method of user portrait dynamic update is proposed,and an intelligent dynamic recommendation system centered on user behavior is constructed.Users’ interests are constantly changing.It is necessary to update users’ portraits in time,grasp users’ new interests,and let new interests immediately feed back to the changes of recommendation results.This paper constructs a dynamic recommendation system for this user behavior.The system mainly includes four parts:feature extraction network,candidate set generation network,dynamic user portrait and ranking learning network.Feature extraction network extracts the content features of user behavior objects in real time to capture users’ interests;The candidate set generation network recalls the candidate set through feature comparison.Aiming at the problem of time consumption in the recall process,PQ inverted index algorithm is adopted to reduce the data dimension and speed up the comparison search and candidate set generation;The dynamic user portrait module realizes the real-time dynamic update of user portrait based on time-varying convolution window;The ranking learning network adopts the Bayesian ranking learning network based on user portrait,takes the user portrait as a priori experience,uses matching to iteratively optimize the ranking optimization strategy,improves the matching degree between the recommendation results and the user portrait,obtains the recommendation probability of candidate set data,and generates the optimal accurate recommendation sequence to realize intelligent dynamic recommendation.The experimental results on the deep fasion data set show that the constructed recommendation system shows good results in accuracy,recall rate,F1 index and the time consumption of candidate set recall.The research results provide practical reference and theoretical support for capturing users’ interests in real time and realizing intelligent dynamic recommendation.
Keywords/Search Tags:social behavior, interest prediction, deep learning, user portrait, intelligent recommendation
PDF Full Text Request
Related items