The Study On The Recommender Algorithm Combining Positive And Negative Feedbacks And Its Application

The application and popularization of Internet technology have brought us to the era of network information. Faced with huge amounts of resource sharing, it is difficult for users to accurately obtain the required information. In order to cope with the problem that it fails to satisfy users for their unique demands in information retrieval, Personalized Recommendation System was born at the right moment and then has been widely used and developed rapidly. It can help users find interesting information. For websites, it is effective to establish relationship between them and users, then to avoid the loss of customers. Therefore, the loyalty of users can be improved. Collaborative filtering is one of the most popular technologies in recommendation systems. Currently, it is also the most widely used and successful approach. The main idea is that it can utilize the history records, like browsing, purchasing, reviewing and so on, and the nearest neighbors with similar hobbies to predict the preference of a certain user.Traditional collaborative filtering algorithms like ItemKNN, can't excavate the relationships between items that are not co-rated by at least one user. In order to solve this problem, the item-based factor models are put forward to utilize low dimensional space to learn implicit correlations between items. However, these models regard all user's rated items equally as positive examples. This is unreasonable and fails to capture the actual preferences of users. To tackle the aforementioned problems, in this thesis, we propose a novel item-based latent factor model, which can consider user's positive and negative feedbacks while learning item-item relationships. In particular, for each user, his rated items are divided into two different sets, i.e., positive examples and negative examples, depending on whether its rating is above the average rating of the certain user or not. In our model, we assume that the predicted rating of an item should be boosted if the item is similar to most of the positive examples. On the contrary, the predicted rating should be diminished if the item is similar to most of the negative examples. A structural equation modeling approach is used to learn two low-dimensional item eigenvectors and then their inner product can take place of the item-item similarity.The main contributions in this thesis are as follows:1?For users, their rated items can be sorted into two categories, i.e., positive examples and negative examples. Correspondingly, the information that positive examples provide is defined the positive feedbacks and what the negative ones reflect is the negative feedbacks. The classification takes full advantage of limited feedback resource and help exactly catching users' interest, which make those item that users prefer rank in the top.2?The eigenvector of items were trained by machine learning. Instead of providing an explicit parameterization for users, we represent users through the items that they prefer. This method bridges the gap of the different categories and enhances the interaction between users and items, strengthens the flexibility and operability of our model. In addition, the item-item similarity can be replaced by an inner product of two eigenvectors and the sum of all the inner products is defined as preference degree for users.3?Efficient integration of implicit feedback. All the feedbacks were also divided into two different sets according to the classification of items, which includes positive feedbacks and negative feedbacks. Correspondingly, the control coefficients were introduced to adjust the proportion of the positive and negative feedbacks in recommendation result.4?An improved recommendation model with amendatory positive and negative feedbacks was proposed in this thesis. Then, we investigated and analyzed the impact of each parameter referred to dimension and preference coefficient separately. And finally we conducted extensive experiments on diverse datasets.To evaluate the performance of our proposed method, we performed comparative experiments in order to verify the effect of classified items. In the end, we discussed the comparison results with other competing algorithms and studied the influence of various parameters of PNSM on the recommendation performance. Comprehensive experiments on two benchmark datasets indicate that our method has significant improvements as compared with existing approaches in both rating prediction and top-N recommendation. It also confirms that the classification of feedbacks is helpful to infer the interest of user efficiently and accurately. As a result, the precision and ranking accuracy were both improved highly in recommender systems.The algorithm in this thesis has been applied to the Juhaokan video recommender system of Hisense smart television. There were about 1 million active users and 4 million play records per day in one of the data platforms. The average play counts had a five percent promotion, increasing to 4.2 million with our algorithm.