Abstract
Social curation is a new trend which has emerged following on the heels ofthe information glut created by user-generated content revolution. Rather
than create new content, social curation allows users to categorise content
created by others, and thereby creating and resharing their personal taxonomies
of the Web. In this dissertation, we collect a large dataset from
Pinterest, arguably the most popular image curation service, and seek to
understand the trend on three levels: content, friends and crowds.
We first take an empirical look at social curation by mining its content
usage. Our data reveals that curation tends to focus on niche items that
may not rank highly in popularity and search rankings. Yet, curated items
exhibit their own skewed popularity, although most users, or curators, act
for personal reasons. At the same time, it also shows that curators with
consistent activity and diversity of interests show more social value in
attracting followers.
This drives us to explore the role of social networks on social curation.
We find that social users are more active and are more likely to
return soon in Pinterest, indicating a bonding effect enabled by social networks.
Then we divide the social network into two subgraphs, according
to whether they are created natively or copied from some other established
social networks (e.g., Facebook) via a social bootstrapping method.
It shows that, when users just join the service, copied network can promote
more social interaction, as it initiates a stronger and denser social
structure than native network. However, social networks are not critical
for information seeking, as a non-trivial number of users’ content are curated
from strangers with high interest matching. In fact, this trend also
holds for social interaction: Users tend to wean from copied friends to
interact more with interest-based native friends over a long-term view.
Finally, we understand social curation as a distributed computation process,
and examine the relationship between curators and crowds. We show
that despite being categorised by individual actions, there is generally
a global agreement in implicitly assigning content into a coarse-grained
global taxonomy of categories, and furthermore, users tend to specialise
in a handful of categories. By exploiting these characteristics, and augmenting
with image-related features drawn from a state-of-the-art deep
convolutional neural network, we develop a cascade of predictors that together
automate a large fraction of curation actions with an end-to-end
accuracy of 0.69 (Accuracy@5 of 0.75).
| Date of Award | 2017 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Nishanth Sastry (Supervisor) |