A robust ranking algorithm to spamming

Ranking problem of web-based rating system has attracted many attentions. A good ranking algorithm should be robust against spammer attack. Here we proposed a correlation based reputation algorithm to solve the ranking problem of such rating systems where user votes some objects with ratings. In this algorithm, reputation of user is iteratively determined by the correlation coefficient between his/her rating vector and the corresponding objects’ weighted average rating vector. Comparing with iterative refinement (IR) and mean score algorithm, results for both artificial and real data indicate that, the present algorithm shows a higher robustness against spammer attack. Introduction. – The abundance of available information troubled people every day, and information filtering technique is quickly developed in recent years. An important aspect in information filtering is the rating system. There are a range of daily examples of rating system. Such systems include opinion websites (Ebay, Amazone, Movielens, Netflix, etc.), where users evaluate objects. Ranking is one of the most common way to describe the evaluation aggregation result, which gives a simple representation of the comparative qualities of objects. PageRank is the most widely applied algorithm for search engines which rank websites based on the directed hyperlink graph [1]. Recently, some iterative algorithms are used in scientific citation network to rank scientists [2]. Both the hyperlink network and scientific citation network are unipartite systems, but many other rating systems have a bipartite structure with two kinds of node: users as evaluators and objects as candidates [3]. In this paper, we consider the ranking problem in such rating systems where users vote objects with ratings, and devise algorithms to accurately rate objects. Ranking objects according to their average ratings is a straightforward statistical method. However, in the open evaluation system, the user can be somebody who are not serious about voting, or he/she is not experienced in the corresponding field and gives some unreasonable ratings. What even worse is that the user might be an evil spam(a)E-mail: [email protected] mer who gives biased ratings on purpose. Therefore, the evaluation by simply averaging all ratings may be less accurate. Building a reputation system for users is a good way to solve this problem [4, 5]. Users with higher reputations are assigned more weight. Such reputation mechanisms are widely used in online systems, such as online shops [6], online auctions [7], Wikipedia [8], P2P sharing networks [9], etc. There are already some ranking algorithms based on reputation estimate [10–13]. In [12, 13], an iterative refinement (IR) algorithm is proposed. A user’s reputation is inversely proportional to the difference between his/her rating vector and the corresponding objects’ weighted average rating vector. Weighted rating of all objects and reputation of all users are recalculated at each step, until the change of weighted ratings is less than a certain threshold between two iteration steps. Kerchove and Dooren [11] modify the iterative refinement algorithm by assigning trust to each individual rating. In most previous works, the influence of spammer attack in rating systems is always ignored. In this paper, we proposed a correlation based ranking algorithm. Reputation of user is determined by the correlation coefficient between the user’s rating vector and the corresponding objects’ weighted average rating vector. By comparing with other algorithms, the effectiveness of the correlation based ranking algorithm was tested using artificial data. The results show that correlation based