Information Technology Reference
In-Depth Information
Table 1.
Simulated interest changes
Learn two topics in parallel Learn a new topic
l.
1
R
6
/R
21
n.
1
R
6
/R
21
→ R
6
/R
21
/R
20
l.
2
R
10
/R
32
n.
2
R
10
/R
32
→ R
10
/R
32
/R
50
l.
3
R
41
/R
79
n.
3
R
41
/R
79
→ R
41
/R
79
/R
58
l.
4
R
26
/R
68
n.
4
R
26
/R
68
→ R
26
/R
68
/R
1
l.
5
R
23
/R
37
n.
5
R
23
/R
37
→ R
23
/R
37
/R
41
l.
6
R
44
/R
53
n.
6
R
44
/R
53
→ R
44
/R
53
/R
79
Forget a topic Penalise a topic
f.
1
R
6
/R
21
/R
20
→ R
6
/R
21
p.
1
R
6
/R
21
/R
20
→ R
6
/R
21
/¬R
20
f.
2
R
10
/R
32
/R
50
→ R
10
/R
32
p.
2
R
10
/R
32
/R
50
→ R
10
/R
32
/¬R
50
f.
3
R
41
/R
79
/R
58
→ R
41
/R
79
p.
3
R
41
/R
79
/R
58
→ R
41
/R
79
/¬R
58
f.
4
R
26
/R
68
/R
1
→ R
26
/R
68
p.
4
R
26
/R
68
/R
1
→ R
26
/R
68
/¬R
1
f.
5
R
23
/R
37
/R
41
→ R
23
/R
37
p.
5
R
23
/R
37
/R
41
→ R
23
/R
37
/¬R
41
f.
6
R
44
/R
53
/R
79
→ R
44
/R
53
p.
6
R
44
/R
53
/R
79
→ R
44
/R
53
/¬R
79
topic
R
3, then we may denote such a change as
R
1
/R
2
/R
3
R
1
/R
2. Similarly,
we present here results for four kinds of simulated interests (or tasks) and six
sets of topics (table 1). In an attempt to overcome known problems with the
large number of test documents per topic in RCV1 [24], we have chosen instead
topics with a small number of relevant documents.
The first task involves virtual users with parallel interest in two topics. It
does not simulate a radical change of interest, but tests the ability of the system
to learn two topics simultaneously and adapt to short-term variations in the
user's level of interest in these topics. In the second task, the initial interest in
two topics is followed by the emergence of a third topic of interest. As already
described in the example, in the third task the virtual user is no longer interested
in one of the initial three topics. The fourth kind of interest change is similar to
the third, with the difference that the virtual user explicitly indicates the change
of interest through negative feedback (denoted with “
→
”).
For each of the above tasks we start with an empty profile that is subsequently
adapted to the initial set of interesting topics. For that purpose we use a set of
documents comprising the first 30 documents per topic in RCV1's training set.
The documents are ordered according to publication date and therefore their
distribution is not homogeneous, but rather reflects the temporal variations in
the publication date of documents about each topic. It simulates fast, short-term
variations in the virtual user's interests. For tasks that include radical changes in
the virtual user's interests (tasks
n
,
f
and
p
), the same process is subsequently
executed using the first 30 training documents per topic in the set following the
change of interest. Training documents that correspond to negated topics have
been used as negative feedback.
During the first adaptation phase for task
l
and the second for tasks
n
,
f
and
p
, the profile is used every five training documents to filter the complete
¬
