Research On Personalized Recommendation Technology In Heterogeneous Social Networks

With the popularity of Internet and the rapid growth of Web 2.0 sites, the infor-mation has an explosive growth which overwhelming people and leading them difficult to find the information matching their interests. Personalized recommending systems have emerged as promising tools to provide personalized services for users by explor-ing their potential interests from the huge amount of information, and have drawn wide attention in academia and industry. Moreover, with the rapid growth of social network-s, recommendation algorithm based on social networks has become a new hotspot of research in the field of recommending systems.As traditional personalized recommending systems mostly focus on the applica-tions with single type of items and single relation between users and items, they can not be applicable in the social network which is a complicated heterogeneous infor-mation network with multiple types of items and multiple relations of users and items. Therefore, how to effectively exploit the information of social networks has become a hot research topic of recommending systems field. Moreover, majority of traditional recommending systems focus on how to improve the performance of recommenda-tion algorithm based on score prediction. However, the true purpose of recommending systems is to find and recommend the top-N objects appealing users, but not to predic-t the score users will rate, so the predicting-score-based recommendation algorithms do not have ranking property. Therefore, the ranking-oriented top-N recommendation algorithm has drawn more and more attention. Besides, compared with implicit us-er feedback like purchases, clicks, explicit user feedback is relatively rare in the real world, while the predicting-score-based recommendation algorithms can only use the explicit user feedback, so they usually have data sparse, cold start and other issues.Based on the existing study of the personalized recommending technologies, we investigate the top-N recommending issues in heterogeneous social networks with im-plicit user feedback. We propose an asynchronized random walk based similarity mea-sure method, HybSim, through analyses on the strength and weakness of existing inter-path based similarity measure methods, which can exploit the intra-path and inter-path similarity between different typed objects. Then we further propose a ranking-oriented top-N recommendation algorithm, HybRec, in heterogeneous social networks with im-plicit user feedback by using HybSim to measure the similarity and using Bayesian Ranking Optimization [44] to qualify and synthesize multiple similarity semantics be-neath meta paths, which can effectively solve the data sparse, cold start and other issues to improve the quality of recommending systems by exploiting the heterogeneous in-formation and implicit user feedback in social networks. The main contributions of our paper are as follows:1. For the weakness of existing similarity measure methods in heterogeneous so-cial networks, we propose an asynchronized bidirectional random walk based similarity measure method, HybSim, which can exploit the inter-path and intra-path similarity semantics within heterogeneous social networks to improve the performance of similarity measure.2. To solve the data sparse, cold start and other issues existing in traditional recom-mendation algorithms, we propose a ranking-oriented top-N recommendation algorithm to exploit the implicit user feedback and heterogeneous information in social networks by using HybSim to measure the similarity and using Bayesian Ranking Optimization to qualify and synthesize multiple similarity semantics.3. We evaluate the performance of HybSim and HybRec with two real-world dataset-s, the dataset from RT-IMDb and the dataset crawled from DouBan website. Empirical studies demonstrate that HybSim and HybRec outperform the state-of-the-art approaches respectively in terms of similarity accuracy and recall rate, which confirm that the quality of recommendation can be improved by synthe-sizing the multidimensional similarities quantified from both the intra-path and inter-path.