Information Technology Reference
In-Depth Information
Fig. 7.3
Two-layer approach to query-dependent ranking
relevance labels. In this regard, this approach can avoid the mismatch between
query clustering/classification with the goal of ranking.
•
In this approach, query features are directly used to adjust the ranking model,
instead of through a finite number of query clusters or categories. In this regard,
this approach is more direct, and can achieve query-dependent ranking at a much
finer granularity than previous work.
Note that this two-layer learning approach is very general and can be applied to
many start-of-the-art learning-to-rank algorithms. In [
6
], He and Liu take RankNet
as an example to demonstrate the use of the approach. Specifically, the original
RankNet algorithm is used for document-layer learning, and a “couple” neural net-
work is created to perform the query-layer learning (denoted as CoupleNet). The
corresponding two-layer learning process can be illustrated using Fig.
7.3
. Experi-
mental results reported in [
6
] show that this two-layer learning approach can signif-
icantly outperform both conventional learning-to-rank methods and previous query-
dependent ranking methods.
7.3 Discussions
In this chapter, we have introduced query-dependent ranking methods. Although
they have shown promising experimental results, we would like to point out that
there are still several important issues to consider along this research direction.
•
Most of the methods rely on query features. The query features are used to per-
form query clustering, classification, and to find nearest neighbors. In the ideal
case, the queries in the same cluster, category, or close to each other, should share
similar ranking models. However, the reality is that the query features might not
be able to reflect the similarity between ranking models. In other words, there is
a kind of “gap” between the information we have (query features) and our goal
(model-based clustering or classification). Note that the two-layer approach can
avoid this problem to a certain degree.
