Research On Personalized Recommendation Based On Optimized Support Vector Machine

In the past decades, the rapidly increasing amount of information on the Internet, people from a lack of information era enter into the “information overload” era. Vast amounts of information make it impossible to quickly and accurately locate the information, which they are interested in, from such a huge information resource. Personalized recommended system is an effective tool to solve this problem, which based on the historical behavior information of the users’ to establish the “user-item” relational model. In this way, it can effectively to mining the users’ interests and preferences, recommended personalized resource for users.Given there are some problems in the personalized recommendation practical application, such as mall sample, nonlinear and high dimension, but the support vector machine(SVM) is just proposed for small sample study, and that in solving nonlinear problems can overcome “curse of dimensionality”, and has a strong advantage in dealing with high-dimensional sparse data problem. So, the personalized recommendation method based on support vector machine is proposed in this paper. Not only to analyze the content information of items, but also to analyze users’ behavior information.The main research work of this article can be summarized as follows:① Against problems in traditional collaborative filtering recommendation, such as the single for calculating the similarity, difficult to use the content information of items, and cold start, etc.. The use of the SVM is proposed to instead of the similarity calculating. Not only consider the users’ behavior information, but also using the content information of items, and demographic information. Meanwhile, to further improve the recommendation accuracy, adopting the proposed particle swarm optimization(PSO) having shrinkage factor with dynamic adaptive inertia weight to optimize the parameters of SVM.② Against practical applications, not only need for the recommended list, but also the detailed ratings information(to some extent reflects the user’s interest degree), the regression method based on SVM classification for recommended method is proposed. Firstly, based on the relationship information of “user-item” to construct feature vectors and train a classification model, thus forecasting the items’ classification and forming an initial recommendation list. Then, build a regression model based on this recommended list to predict the items’ scores. At the same time, the proposed PSO algorithm with adaptive evolution speed and aggregation is adopted to optimize the parameters of prediction model.③ Considering the recommendation efficiency and real-time requirements of dealing with the large-scale data, the recommended method of twin support vector machine(TWSVM) based on smoothing techniques and kernel reduced techniques is proposed. The method uses smoothing techniques to transform the TWSVM, avoiding large-scale matrix inverse operation, reducing the time complexity. To enhance the ability of dealing with large-scale data, adopting kernel reduced techniques to further reduce the time complexity and space complexity. Meanwhile, in view of the user’s interests and preferences will be evolving with the time, place and other factors, so the real-time requirements for recommendation system is higher. For this reason, introducing a feedback mechanism, add the user’s rating data to the historical data set in time, and design training rules, start re-training model, so that the model has a certain adaptive ability, to improve the recommendation quality.④ Against the high value labeled data but scarce and labeling the unlabeled data existing time-consuming, labor-intensive, high-cost. Therefore, the recommended method based on semi-supervised learning combing transductive support vector machine(TSVM) with active learning(AL) is proposed. Firstly, mining the valuable user comments and adding them to relationship data set. Then, using the batch mode AL strategy to query and label the unlabeled data which has the highest information from the large number of unlabeled samples set, obtaining as small as possible and the most valuable sample set for enhancing the classifier, thereby reducing the cost of labeling sample, improving the classification performance. Meanwhile, in order to better utilize the distribution information of unlabeled data, adding a graph-based manifold regularization term to objective function to further enhance the effect of the recommended model.