Information Technology Reference
In-Depth Information
Fig. 4.10
The mMAP for the examined methods:
a
personalized image search,
b
topic-based image
recommendation
contact users, it is weird that the performance of
Social_global
is inferior to that of
Basic_AC
. The reason may be that
Social_global
ignores the relation to the query
and treats the social influence as constant for all topics. This result coincides with the
motivation illustrated by the toy example we introduce at the beginning of the chapter.
By considering the topic-sensitive influence values for different queries,
Social_topic
obtain better performance than both
Basic_AC
and
Social_global
, which validates
the advantage of the proposed framework to incorporate topic-sensitive influence
modeling for personalized search.
4.5.3.2 Topic-Based Image Recommendation
In social media web sites, users will explicitly fill in interest profile or express their
interests by following certain themes, e.g., sport, music, travel, etc. In this case,
users will appreciate if resources based on the expressed user preferences are recom-
mended. In this chapter, we simulate the process of expressing interests by assuming
user selecting one extracted topic and evaluate the performance of topic-based image
recommendation.
The
Favorite
images are again employed as ground-truth for evaluation. The
intuition is that, if image
d
j
is marked
Favorite
by user
U
i
and its dominant topic
is
Z
t
, then
d
j
is treated as relevant when
U
i
selects topic
Z
t
. In our experiments,
100 randomly selected users who marked
Favorite
to 50-100 images constitute
the test set and a total of 7,725
Favorite
images constitute the relevant image set
for the topic-based test queries. Following the proposed risk minimization-based
framework in Eq. (
4.16
), query
q
is replaced by the selected topic
Z
t
and we can
compare between the three topic-sensitive influence modeling methods: TAP, mTIH,
and mmTIM. mMAP is utilized for evaluation and the results are demonstrated
in Fig.
4.10
b. It is shown that the proposed mmTIM obtains better performance
than the other two methods, which is consistent with the results shown in Fig.
4.9
.
This further validates the effectiveness of incorporating multimodal information for
image-related tasks. The fact that mmTIMandmTIH outperformTAP coincides with
out motivation that more accurate topic-sensitive influence modeling contributes to
better recommendation and search results.
