Link Prediction Based Recommendation Algorithm Research In Complex Network

With the rapid development of computer, internet and web technology, people’s lives are undergoing ever-lasting revolutions in all aspects. People are accustomed to online exchange thoughts, make new friends in virtual communities, browse webpages, read news and books, watch movies and shop onlin, basically differring from the original lives. They experience the abundant and significant lives derived from the fast development of internet and meanwhile endure severe explosion of an ocean of information and data, resulting in new-coming annoyance, because people can not rapidly find out the most related information in an ocean of data consisting of thousands of movies, millions of books and billions of web pages. The sharp increase of data has exceeded the people’s information processing ability, leading to an unimaginable cost of information retrival. Manual evaluation and choice on goods becomes more and more difficult, bringing in a dilemma in which urgent people can not find their needed commodities in extremely rich commodity collection repository. However, fortunately, the most attractive and outstanding recommendation system exhibits excellent information filtering ability. It discovers users’ personal habbits according to their acitivity history records and recommends them commodity objects based on their discovered personal habbits. Nowadays, the applications based on recommendation technology spread everywhere, such as Amazon.com recommending books to customers according to their books buying records, AdaptiveInfo.com recommending news to readers according to their reading history and TiVo digital TV system recommending TV programs and movies according customers’watching patterns and rating scores.Among various kinds of recommendation algorithms, link prediction based recommendation algorithm has attracted a lot of attention. Therefore, beginning with study of link prediction in network of unique kind of node, this paper further applies link prediction study to bipartite network to find similarity of objects, ultimately complishing collaborative recommendation. Highlights list as follows:1. By considering weak tie on unweighted network, we propose local path based modified link prediction algorithm. Traditional similarity based algorithms, especially AA (Adamic Adar) and RA (Resource Allocation), ignore the impact on nodes’similarity from the neighbor’s weak relationship on unweighted network, leading to a limitation on prediction performance. Therefore, this paper based on weak relationship, also called weak tie, proposes improved local similarity link prediction algorithms OAA (Optimized AA) and ORA (Optimized RA), emphasizing the weak ties and significantly enhance accuracy and adaptation.2. Upon finding heterogeneity of paths, we propose SP (Significant Path) link prediction algorithm. It is found that in local similarly based link prediction, paths with different structures have different abilities in passing similarity between two endpoints, and especially the paths consisting of nodes with small degrees can pass more similarity. Meanwhile comparatively longer paths can increase more pipes for passing similarity. Therefore, this paper proposes link prediction algorithm SP (Signifiant Path) based on path heterogeneity, assigning different paths with different weights to enhancing accuracy of link prediction.3. We propose novel link prediction algorithm by considering effective influence of endpoints. It is found that traditional link prediction algorithms neglect the non-contribution relations in endpoints and overestimate the influence of endpoints, wrongly enhancing similarity between endpoints and weakening accuracy of link prediction. Accordingly, this paper proposes a novel link prediction algorithm named as EP (Effective Path), eliminating non-contribution relations and extracting effective influence. Moreover, EP emphasizes the paths consisting of nodes with small degrees, ultimately in real data experiments obviousely outperforming traditional algorithms in accuracy.4. In terms of backward similarity from uncollected objects to collected objects, we proposed corrected similarity index in recommendation. It is found that under the studies of resource diffusion on in network and topology properties in bipartite network, link prediction algorithms can be used for similarity based recommendation. However, because of sparsity and asymmetry of network, the similarity estimation will be overestimated or underestimated. So under the consideration of the backward similarity from uncollected objects to collected objectsd, this paper corrects the single directional similarity estimation, conquering sparsity of network, proposing an improved recommendation algorithm named CSI (Corrected Similarity Inference). Moreover, experiments’results show CSI obtains remarkable accuracy, diversity and novelty in contrast to traditional algorithms.5. We propose the unbiased consistent recommendation algorithm. It is found that traditional algorithms recommending uncollected objected according to the temporally causality, but in most cases the purchasing sequences do not mean any causality. Actually, the intrinsic reason for perchasing two objects exists in consistent preference of the two objects. Therefore, this paper proposes CBI (Consistence-based Inference) and UCBI (Unbalanced CBI) recommendation algorithms, obtaining significant improments in accuracy, diversity and novelty in real data experiments.