Information Technology Reference
In-Depth Information
Table 12.1
Dataset 1 for
Yahoo! learning-to-rank
challenge
Training
Validation
Test
Number of queries
19
,
944
2
,
994
6
,
983
Number of URLs
473
,
134
71
,
083
165
,
660
Table 12.2
Dataset 2 for
Yahoo! learning-to-rank
challenge
Training
Validation
Test
Number of queries
1
,
266
1
,
266
3
,
798
Number of URLs
34
,
815
34
,
881
103
,
174
The competition is divided into two tracks:
•
A standard learning-to-rank track, using only the larger dataset.
•
A transfer learning track, where the goal is to leverage the training set from
dataset 1 to build a better ranking function on dataset 2.
Two measures are used for the evaluation of the competition: NDCG [
2
] and
Expected Reciprocal Rank (ERR). The definition of ERR is given as follows:
1
,
m
i
−
1
1
m
G(y
i
)
16
G(y
j
)
16
2
y
ERR
=
−
with
G(y)
=
−
1
.
(12.1)
i
=
1
j
=
1
The datasets can be downloaded from the sandbox of Yahoo! Research.
1
There
are no official baselines on these datasets, however, most of the winners of the com-
petition have published the details of their algorithms in the workshop proceedings.
They can serve as meaningful baselines.
12.2 Microsoft Learning-to-Rank Datasets
Microsoft Research Asia released two large-scale datasets for the research on learn-
ing to rank in May 2010: MSLR-WEB30k and MSLR-WEB10K. MSLR-WEB30K
actually has 31,531 queries and 3,771,126 documents. Up to the writing of this topic,
the MSLR-WEB30k dataset is the largest publicly-available dataset for the research
on learning to rank. MSLR-WEB10K is a random sample of MSLR-WEB30K,
which has 10,000 queries and 1,200,193 documents.
In the two datasets, queries and URLs are represented by IDs, with a similar
reason to the non-disclosure of queries and URLs in the Yahoo! Learning-to-Rank
Challenge datasets. The Microsoft datasets consist of feature vectors extracted from
query-URL pairs along with relevance judgments:
1
http://webscope.sandbox.yahoo.com/
.