“Yahoo发起了一项学习排序竞赛(Learning to Rank Challenge)作为ICML 2010大会的一部分，任何人可以以个人名义或组队（最多10人）参赛。竞赛3月1日开始，至5月31日结束，6月份公布获奖名单。
竞赛将公布两个之前从未发布的真实数据形成的数据集。第一个数据集包括29921个请求，744692个URL地址，519个特征。第二个数据集包括6330个请求，172870个URL地址，596个特征。竞赛的任务是根据训练集中的数据构造一个排序函数，对验证集和测试集中URL地址进行排序。
第1至4名优胜者将分别获得8000，4000，2000，1000美元，并将被邀请参加ICML 2010大会。
更多详细信息参见：Learning to Rank Challenge“
The task of learning to rank has recently drawn a lot of interest in
machine learning. As distinguished by [3] and [4], previous works
fall into three categories: (1) the point-wise approach, (2) the pairwise
approach, and (3) the list-wise approach.
In the point-wise approaches, each training instance is associated
with a rating. The learning is to find a model that can map
instances into ratings that are close to their true ones. A typical
example is PRank [5], which trains a Perceptron model to directly
maintain a totally-ordered set via projections. The pair-wise
approaches take pairs of objects and their relative preferences as
training instances and attempt learning to classify each object pair
into correctly-ranked or incorrectly-ranked. Indeed, most existing
methods are the pair-wise approaches, including Ranking SVM
[10], RankBoost [9], and RankNet [2]. Ranking SVM employs
support vector machine (SVM) to classify object pairs in
consideration of large margin rank boundaries. RankBoost
conducts Boosting to find a combined ranking which minimizes
the number of misordered pairs of objects. RankNet defines Cross
Entropy as a probabilistic cost function on object pairs and uses a
neural network model to optimize the cost function. Finally, the
list-wise approaches use a list of ranked objects as training
instances and learn to predict the list of objects. For example,
ListNet [3] introduces a probabilistic-based list-wise loss function
for learning. Neural network and gradient descent are employed
to train a list prediction model.
—–J. Yeh, J. Lin, H. Ke, and W. Yang. Learning to rank
for information retrieval using genetic programming.
In LR4IR, 2007.
When ranking problem is described as a machine
learning problem, proposing and minimizing the ranking
loss function becomes the key to learning to rank. There
are several popular approaches to constructing the
ranking loss function which are considered on different
instance level. One is building the loss function on
document instance level. SVOR [9] is proposed to
minimize the rank loss by aggregating the error on each
document instance. Another approach is pair-wise loss
function, which create the pair instance between two
documents with different relevance level, and denote
correct rank pair as positive (+1) instance while incorrect
as negative (-1). So the ranking problem is transformed
into a binary classification problem in RSVM [1],
RankBoost [2], RankNet [3]. And a recent approach is
list-wise, which define rank loss with the difference
between predicting document list and labeled list for each
query, in AdaRank [10], ListNet [11].
—–An Ensemble Approach to Learning to Rank