Research On The Key Technologies Of Preference Prediction For Internet Users

With the development of the Internet and mobile computing technology,Internet applications,such as social network and e-commerce are rapidly spreading.The speed and scale of the data generated in the network is unprecedentedly high.The modern society is stepping into the period of big data.A lot of information related to users' preference are contained in the data.We need to seek out the contents that grab users' eyes from amounts of information and to precede users' preference demands us to be fast and accurate.We also need to figure out the value of these information and finally make it work to serve the users and produce personalized recommendations.All of these requirements are hot pots in Internet Data Science area.The empirical Internet world does own some characters as too many hosts and enormous datas without focus.Traditional methods couldn' t match that any longer.These methods mainly based on manual work,statistics and experimental fomulas,singly Collaborative Fliterings,and so on.What is more,the behavior of network users' is appearing as more multiform and tanglesome.When we came to a situation what calls us to figure out the essential rules of users' exactly and fix the position of users' preference,the traditional analysis model is not useful any more.Besides,new solutions are provided by machine learning and AI developing to analyze users' preference,but they are still in the exploratory stage.In consideration of this,the passage is aimed at personalized recommendations.This paper will put most attentions on the problems that present in the progress of analyzing users' preferences,such as incomplete preference expressions,imprecise models and the influences of special elements.It's a new solution that we can study and discuss the pivotal technology in users' preference analyzing.In that,this paper start from a thinking of exercising machine learning.By considering the elevation of model's definition,the advantage of recommending rate,we will focus on the key technologies of user preference prediction.Our research will follow the following orders: the order of acquiring user preference id from direct to indirect,the order of research sence is from general to special.furthermore,rating prediction,ranking optimization based on user preference,influence of side information,preference predictions for special application scenarios will be the mian points of view in this paper.The work of the dissertation is supported by the National Natural Science Foundation of China under Grant 61271 308 and 61172072,Beijing Natural Science Foundation underGrant 4112045 and the Fundamental Research Funds for the Central Universities2016YJS028.Main contributions of the dissertation are as follows:1.By studying the influence that static parameter made to prediction accuracy of rating preference,we put forward a predicting model of rating preference changing with dynamic parameter.User rating is a kind of data about users' preference.It can display users' preference visually and be made use of directly.The passage mainly studies the matrix decomposition model of rating prediction.There are two important questions appearing in matrix decomposition model,fixing regularization parameter and the distribution of predictive sore with real rating rule unmatched.To solve them,the passage considers it will be useful to improve the model of dynamic matrix parameter and fine tuning.Two innovative thinkings are showed in the passage as following: Firstly,on the base of foll researching on changing rule of prediction accuracy during the matrix factorization progress,we put forward a parameter measurement arrithmetic about dynamic regularization.The arithmetic redefines regularized parameter,instead of a fixed value.The regularized parameter can accommodate the optimizing progress of matrix decomposition and change dynamically by itself.Secondly,we put forward a fine tuning arithmetic of rating aimed at the final prediction matrix in order to solve the gap of rating distribution between the initial matrix and predictive matrix.This arithmetic can obtain a best predictive result among the whole progress.What is more,it can also adapt to the real rule of rating distribution.There comes to the conclusion from the experiment that,comparing to the traditional matrix distribution model,the new one can improve the prediction accuracy of rating predicting,and make the prediction fits the rating distribution rule of real datas.2.We bring forward a self-adaption updating model of users' preference vector which based on studying rating of preference vector.The model is advanced to figure out the imprecise existing in initial preferring vector.Except for scores,users' preference vector is also one of methods that be frequently-used to express users' preference directly.But it is not intact and accurate enough on expression.Considering of that,we put forward a improvement and a brand model based on the characters of content composition and the expressing style of users' preference vector.We discussed the traditional self-adaption updating model based on preference studying analytically.The new model is designed to the disadv 1antage of the old one and to adapt a actual application environment.The innovations of thisnew model are represented as following: Firstly,according to improving accuracy of it,the model observes and collects mainly on implicit feedback generated interactively by users and recommendation system.A strength arithmetic influenced by candidate of overflow area is brought forward to resolve unbalanced influence about the overflow area of the implicit feedback options occurring in selfadaption updates.Besides,the weighted management mechanism is introduced in the model.Secondly,according to actual application,we advance a quantized arithmetic based on the users' multi-attribute of sliding windows in order to figure out the real choice of users' during the progress of making decisions actually.For this consideration,the model is aimed to explain the phenomena that different users pay attentions in different degrees on multi-attribute of recommendations in actual application environment.The analyzation of datas shows that the performance of recommendation ranking can be great promoted.The performance is generated by the preference vector after self-adaption updating in actual recommending scene.3.By studying the influence of field factor in predicting user preference,we propose a field-aware model to predict user preference.In real circumstances,direct feedback,such as rating,that we can collect from users is very rare.Meanwhile,only limited information can be mined by traditional user preference prediction models,because they simply consider user-item two dimensions.To cure the above problems,this paper extends the study dimension to "field",focus on studying the impact of analyzing user preference generated by side information(such as field).This paper clearly discuesses the definition of field in user preference analysis,and explains why it is important to add field factors to guide user preference prediction.An improved method is proposed to establish a new model of preference prediction,the model is inspired by existing methods and it is proposed to against the disadvantages.The main innovation points of this paper are as follows: the model directly map field information into the latent space of the framework of matrix factorization,it will learn three kinds of potential relationships,which is relations between user and item,telations between user and field,relations between item and field.The proposed model is used to solve the problem of top-N recommendation.Empirical data analysis shows that the proposed model has a significant improvement in the top-N recommendation accuracy.4.By studying the characteristics of location-based social network,for the influence of geographical and social factors on the prediction accuracy of position preference,alocation prediction model is proposed.This paper focuses on the special application scenario of location-based social network,make full use of location data as side information to study users' location preference.To improve the prediction accuracy of location preference,we analyzes three factors that influence location preference:content,social and location.The main innovation points of this paper are as follows:for content impact factors,we propose multi-location-tag extraction method,by using this method,a user-item matrix can be translated into a user-tag matrix for subsequent analysis;for social impact factors,we present a method of interest similarity measurement between friends;for location impact factors,we propose a location influence quantization method based on distance attenuation,this method can model the influence of the distance between the location and the user in real environment.The experimental analysis shows that the proposed model conforms to the real location-based social environment and can effectively improve the efficiency of location recommendation.