Research On Personalized Recommendation Based On Probabilistic Relational Models

In recent years, with the speed development of internet and e-commerce techniques, the number of information and types of commodities offered by the enterprises increases rapidly, personalized recommendation system provides an efficient method in order to solve the information overload problem. It simulates shop sales staff to recommend personalized products to users, helps users find the necessary goods and makes them have a successful purchase process. Personalized recommendation system for e-commerce has gradually become an important IT technology research, receives more and more attentions from the researchers.The research of this paper is funded by the National Natural Science Foundation named"Research on A Number of Issues in Statistical Relational Learning"and"Internet-based Network Educational Resource Management System, NERMS", which is a scientific and technological development project in Jilin Province. NERMS system's main target is the effective range of the organization and management to numerous of the network educational resource, in order to efficiently share, conveniently obtain the network educational resource, speed up the exploitation of the network educational resource and promote the development of network education. In order to achieve this goal, NERMS is designed as a Website focused on intelligence and personality. We have made a series of research work on intelligence and personality of the Website, and the personalized recommendation technology based on Statistical Relational Learning is one part of them.This paper analyzes and summarizes the personalized recommendation system and the Probabilistic Relational Models theory in Statistical Relational Learning; applies PRM into personalized recommendation system, proposes a unified recommendation model based on PRM. The model includes customer-based, project-based, collaborative filtering and hybrid models. The hybrid model makes the other three models be weighted and switched. Then the paper verifies the four models'performance in the testing datasets; applies the recommendation models based on PRM into NERMS, designs and implements personalized recommendation subsystem, which provides the Website users with initiative resource recommendation. This paper mainly includes three aspects as follows.Firstly, the paper analyzes and summarizes the personalized recommendation system and the Probabilistic Relational Models theory. Personalized recommendation system is a procedure that recommends information and commodities according to users'personal preferences and habits. Its workflow can be divided into three parts which are input (information gathering) function modules, recommended method (data processing) modules and output (recommendation results displaying) functional modules. The information pretreatment technology of recommendation method includes the drop-dimensional of user's evaluation, the densification and standardization; Researches on some mature recommendation algorithms based on association rules, clustering, the Nearest Neighbor, collaborative filtering of Rating Prediction, Bayesian network technology and Horting map etc. Subsequently, indicates the problems that the personalized recommendation technique encountered such as sparsity, scalability, real-time, accuracy, Cold-start, singular found, and so on. Probabilistic Relational Model extends the standard attribute-based Bayesian network model (Bayesian) which uses the Entity Relationship (ER) model as the basic framework and described as a template for a probability distribution over a set of instances of a given schema. The template includes a relational component, that describes the relation schema for the domain, and a probabilistic component, that describes the probabilistic dependencies that hold in the domain. There are two variants of the learning task: parameter estimation and structure learning. In the parameter estimation task, we assume that the qualitative dependency structure of the PRM is known. The learning task is only to fill in the parameters that define the CPDs of the attributes. In the structure learning task, there is no additional required input. The goal is to extract an entire PRM, structure as well as parameters, from the training database alone. This paper researches on the maximum likelihood estimation on complete relational data and the heuristic search algorithm based on greedy hill-climbing searching.Secondly, the paper applies Probabilistic Relational Models into personalized recommendation system, and proposes a unified personalized recommendation model based on PRM. The model utilizes project information, user information and ratings matrix to construct the PRM. According to the PRM dependency structure and parameters estimation, namely the parents'nodes information and the CPDs of the rating node, the recommendation algorithm predicts the project ratings to which the target user has not rated. And then recommends the projects that win the highest rating to target users to complete the process of the personalized recommendation for E-Commerce. This paper proposes three PRMs to personalized recommendation system based on customer characteristics, project characteristics and collaborative filtering technology, which are called PRM_ITEM,PRM_USER,PRM_CF respectively. Each model has two components: parameter learning algorithm and recommending algorithm. The recommendation model based on PRM can make itself not depend on particular user or project with its first-order characteristic. Once the PRM has been constructed, the recommendation model is highly efficient and can achieve the real-time requirements. Among them, PRM_ITEM can well make use of project characteristics while PRM_USER uses characteristics vectors with the rating matrix to the project, and PRM_CF only utilizes ratings matrix to construct PRM. This paper also proposed a combined recommendation model called PRM_HYBRID, which is obtained through weighting and switching treatments of the three recommendation model mentioned above. PRM_HYBRID can not only satisfy the interest of target users, with consideration of the users'neighbors'habits, but also mine the potential interests of the target user. At the same time, the accuracy and computational complexity of PRM_HYBRID has achieved a better balance. Subsequently, on the dataset of MovieLens, uses 5-fold cross validation to verify the recommendation performance of our four models. As a result, under the two testing datasets which are divided by users'ratings matrix uniformly and randomly, the mean absolute error (MAE) which we adapt as an ideal evaluation for recommendation to assess the quality of the algorithms, is relatively low and stable.Thirdly, this paper applies the personalized recommendation models based on PRM to NERMS. First the author briefly introduces NERMS system's main functional modules, as well as software architecture, and then describes the personalized recommendation subsystem implementation steps, including information obtaining, information preprocessing, making recommendations and results displaying. Access to information obtained is an implicit method, using Website logs which contain the users'actions such as browsing, collecting and downloading the resource. Information processing is to process the data obtained above into the data format that the recommendation algorithms need, which contains user information, resources information and users'evaluation information, and then adjust users'interesting on different levels of the corresponding weighting. In the making recommendations phrases, the administrator calls the recommendation algorithms to make recommendations for the users who have enrolled. In the results displaying phrases, the system will return the recommendation results to users by showing them on the personalized pages in the form of Top-N lists. At present, NERMS has been put into use, after a long period of testing, with stable operation, good performance and the personalized recommendation results are also satisfactory.The personalized recommendation model based on PRM mentioned in this paper has a theoretical significance and practical value, as well as certain reference value when applying PRM into recommendation system. In addition, there are some defects in the models, such as the complexity of the structure learning algorithm, no consideration of incomplete data in parameter estimation and so on. These problems still need to be improved in the future.