The Web has become an ubiquitous environment for human interaction, communication,
and data sharing. As a result, large amounts of data are produced. This
data can be utilised by building predictive models of user behaviour in order to support
business decisions. However, the fast pace of modern businesses is creating the
pressure on industry to provide faster and better decisions. This thesis addresses
this challenge by proposing a novel methodology for an effcient prediction of user
behaviour. The problems concerned are: (i) modelling user behaviour on the Web,
(ii) choosing and extracting features from data generated by user behaviour, and
(iii) choosing a Machine Learning (ML) set-up for an effcient prediction.
First, a novel Time-Varying Attributed Graph (TVAG) is introduced and
then a TVAG-based model for modelling user behaviour on the Web is proposed.
TVAGs capture temporal properties of user behaviour by their time varying component
of features of the graph nodes and edges. Second, the proposed model allows
to extract features for further ML predictions. However, extracting the features and
building the model may be unacceptably hard and long process. Thus, a guideline
for an effcient feature extraction from the TVAG-based model is proposed. Third,
a method for choosing a ML set-up to build an accurate and fast predictive model
is proposed and evaluated. Finally, a deep learning architecture for predicting user
behaviour on the Web is proposed and evaluated.
To sum up, the main contribution to knowledge of this work is in developing
the methodology for fast and effcient predictions of user behaviour on the Web.
The methodology is evaluated on datasets from a few Web platforms, namely Stack
Exchange, Twitter, and Facebook.
|Date of Award||2017|