Research On Recommendation Based On The Trend Of Interests From High Performance Users

The information obtained by users is explosively growing due to the rapid development of the Internet and e-commerce, which brought the rich products and convenience to the users, while users are plagued with problems such as massive data, species diversity, and difficulty in distinguishing truth and fause, which forms a so-called "information overload" problem. Recommender systems are applied to solve this issue. Unlike search engines and web portals, the recommender system can analyze the users dataset, e.g. the partnership amoung users, purchasing history, etc., and predict the products that users have not bought yet but suit their interests. As a result, recommender systems can transform users from hesitation to purchase, which effectively promote users’ trust towards the websites and improve the sales of goods. However, the growth of websites size and number of users has lead to the difficulty of capturing the trend of interests from each user, which is reflected in three aspects. First, the number of users who actively enroll in the interactions with the system is relatively small, and the recommender system could hardly know the real needs of the users. Second, users’trust towards other users and the recommender system is very low, which causes the extraction of valuable users’data and prevents the understanding of users’personalized needs and enhancing the satisfaction towards the system. Third, recommender system has to access users’ personal data to generate accurate recommendations, which is in conflict with the high level of privacy concern from users, and data collection may cause users’ privacy disclosures become conservative and privacy collection strategies in recommender system cannot fully consider the privacy sharing preference of all users. These problems are not conducive to the recommender system to fully understand users’individual needs, which further reduce the recommendation accuracy and users’satisfactions.This dissertation focuses the above three aspects as research objectives, improves the machine learning techniques to mining the trend of interests from high performance users who have high degree of activity, high degree of trustworthiness, and high volume of privacy disclosures, to generate personalized recommendations with high accuracy, and thus make users present trust and satisfaction towards recommender system in e-commerce websites. The main contributions of this dissertation are presented as follows:1. A novel recommender algorithm Div-clustering is proposed in this dissertation, which modeled the data structure of users in e-commerce, and improved the k-means clustering techniques to recognize the users of high activities. The data of users of high activities are applied to promote recommended items, which improve the prediction accuracy of the recommender systems. Div-clustering first analyzes the data structure of users and items, construct the graph model consisting users’interactions and partnerships, and studies the characteristics and traits of highly active users and recommended items. In the experiment section, the dataset of scholarly papers was crawled by Websphinx from Elsevier and IEEE, and movie dataset were downloaded from MovieLens and IMDB. The experiment had shown that our Div-clustering recommender algorithm could provide higher prediction accuracy and stability both in online test and in offline evaluations compared with the traditional recommender system which had not used the users of high activity, and also certificate that the recommendation generated from computing the similarity from the users of high activity could more likely be accepted and trust than from computing the similarity from the normal users.2. Aiming at the problem of data sparsity, a novel recommendation algorithm PointBurst is proposed to mine the latent trust partnerships among users based on the fact that the recommendations from users of high activity are more likely to be accepted and trust. The mined latent trust partnerships provides a powerful supplement to the explicit trust partnerships among users, and eases the problem of data sparsity faced by traditional recommendation methods such as collaborative filtering recommendation. PointBurst mainly analyzes the trust relationship and strength between users on the base of the graph model, optimizes the classifications and connections of users and items, recognizes the users of high degree of trustworthiness and extracts the latent trust relationship between users, and apply the explicit and latent trust relationship to the input of the recommender system for the generation of recommendations. The performance of PointBurst and a baseline recommendation algorithm was compared on the dataset of del.icio.us, Myspace and MovieLens, and shows that the recommendations from PointBurst are more accurate and stable due to its utilization of extracted latent trust relationship from the users of high trustworthiness.3. Based on the previous finding that the recommendation generated from the data of users in high degree of trustworthiness are more likely to be trust, a novel learning model ISBP is proposed, which exploring the latent factors that affects users’ decisions on sharing their privacy with the trust relationship in recommender system, so as to recognize the users of high volume in privacy disclosure and make the recommender system to know the users better for recommending items in high accuracy. After gathering the factors that may affect privacy disclosure from the related works, the ISBP model has construct a group of hypotheses, and improves decision tree classifier, KNN classifier, and naive Nayes classifier to be able to discover the factors are truly influencing the privacy disclosure. After running our ISBP model on the dataset that we collected from crowdsourcing platform SOJUMP, we have found that younger users and users who do not majored in computer have the highest volume on privacy disclosure, while users’ gender does not make the difference. Under the premise of not raizing all users’ privacy concern, the recognized users of high volume in privacy disclosure can share more personal information than other users, which can help the recommender system know them better and generate recommendations with higher accuracy.4. We have discovered a spill-over effect which causes the loss of users in high volume of privacy disclosure, and to ease this effect a decision support strategy DSS is proposed. The spill-over effect means a bad data collection will make the users who used to disclose high volume of privacy tend to disclose less in the resume requests, which further leads to the recommender system to know less about the users and reduce the recommendation accuracy. The spill-over effect is existing in most of the datasets in our previous study, which is mainly caused by high sensitive requests and the users in high volume of privacy disclosure are the major victims. We have modified the clustering techniques for detecting the psychological congnitive factors change during users’ privacy disclosures are less, and found that this cognitive ability could lead to the currency of spill-over effect, while the foundamental reason is that users have no enough background knowledge to support their decisions on privacy disclosure. DSS is proposed to support users make decisions such as help users understand how the recommender system works and gain their background knowledge. The experiment has certificated that DSS can ease the negative effect of the spill-over effect, while maintain the users in high volume of privacy disclosure to supplement enough information to the recommender system and keep the recommendation accuracy high.