Abstract
This thesis shed light on the Internet-based social network link predictionproblem. After reviewing recent research achievements in
this area, two hypotheses are introduced: (i) The performance of
topology- based network prediction methods and the characteristics
of the networks are correlated. (ii) As networks are dynamic,
the performance of prediction can be improved by providing different
treatment to dierent nodes and links. To verify the Hypothesis
(i), we conduct experiments with six selected online social
networks. The correlation coecients are calculated between six
common network metrics and ten widely used topology-based network
link prediction methods. The results show a strong correlation
between Gini Coecient and Preferential Attachment method.
This study also reveals two types of networks: prediction-friendly
network, for which most of the selected prediction methods perform
well with an AUC result above 0.8, and prediction unfriendly
network that on the contrary. For Hypothesis (ii), we proposed
two network prediction models, the Hybrid Prediction Model and
Community Bridge Boosting Prediction Model (CBBPM). The hybrid
prediction model assumes network links are formed following
dierent rules. The model linearly combines eight link prediction
methods and the evolvement rules have been probed by nding
the best weight for each of the method by solving the linear optimization
problem. This experiment result shows an improvement
of prediction accuracy. This model takes link prediction as a time
series problem. Dierent from Hybrid Prediction Model, CBBPM
provides a dierent treatment on nodes. We dene and classify
network nodes as community bridge node in a novel approached
based on their degree and links position in network communities.
The similarity score that calculated from the selected prediction
methods is then boosted for predicting new links. The results from
this model also show an enhancement of prediction accuracy. The
two hypotheses are validated using the research experiments.
| Date of Award | 2017 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Colin Cooper (Supervisor), Sophia Tsoka (Supervisor) & Peter McBurney (Supervisor) |