Learning to Rank: A Machine Learning Approach to Static Ranking

Over the last decade, the Web has grown exponentially in size. Unfortunately, the number of incorrect, spamming, and malicious sites has also grown rapidly. Despite of that, users continue to rely on the search engines to separate the good from the bad, and rank results in such a way the best pages are suggested first. The probably most prominent ranking methods is PageRank [4]. Although Google, the most popular search engine, ranking algorithm originally based on the PageRank algorithm, there has recently been work showing that PageRank may not perform any better than other simple measures on certain tasks. For example Amento et al. [1] found that simple features, such as the number of pages on a site, performed as well as PageRank. We will introduce and discuss two solutions, RankNet [2] and fRank [5], which follow a line of research completely different from PageRank, Hits [3] or other similar ranking algorithms which are based on link analysis. In [2] Burges et al. proposed RankNet, a probabilistic cost for training systems to learn ranking functions using pairs of training examples. In [5] Brill et al. combined all kinds of Web page’s features using RankNet to achieve a ranking system, called fRank (feature Rank), that according to their experiments, is significantly better than PageRank.