Research On Loss Function Of Learning To Rank

With the breakneck-speed of information increment in this day and age, information retrieval technology is becoming crucial important. There are mainly two categories of traditional information retrieval models, namely query-dependent models based on content, such as TF/IDF, probability model and language model and query-independent models based on link analysis, such as PageRank and HITTS. The core of both the two categories is ranking. With the advent of Learning to Rank technology, it is easy to incorporate different information retrieval models into one super-power model, which leads performance improvement.Learning to rank is the cross field of information retrieval and machine learning. Based on how they treat sets of ratings, they can be categorized into the following three groups: pointwise, pairwise and listwise approaches. Loss function is used as the measurement of loss generated by ranking function in the process of training, and thus it plays a pivotal role of learning to rank.This thesis studies the loss function of learning to rank. Firstly, the weakness of adopting different input instance (i.e. sampled in different input space) exclusively is analyzed. Pairwise approach is used as an example to illustrate the improved approach, that is, incorporating the pointwise loss with the pairwise loss function in order to enrich the objective loss function, in which way the real loss in training process is better measured.Secondly, a new listwise approach is proposed based on query-level regression, whose ranking function is modeled by neural network and optimization is carried on by gradient descent.Finally, a framework of incorporation of loss functions is proposed and tested, based on which three weighting schemes for incorporation are given and tested. Further, the three methods using the different merging strategies are compared with the pointwise, pairwise, listwise and other similar approaches that also take into consideration of multiple input instances. This brings a new idea for improving the learning to rank approaches.The experimental result on LETOR demonstrates that the approaches proposed in this paper outperform the existing learning to rank methods.