Abstract:
"
Introduction : Modern recommendation systems are built by a combination of content based
and collaborative filtering methods. Although sometimes the quality of the recommendations
given is somewhat lacking. This is mainly due to two reasons, the misuse of certain movie
metadata and automated systems that forgo any personalization or transparency. In this
dissertation a system is built that makes up for these issues by making use of a movies most
underused attribute, it’s plot and combining it alongside it’s other attributes as well as allowing
the user to select the bias filters on which the recommendations are based on.
Method : The NLP technique of topic modeling was used to breakdown the movie plot and
subject it to a similarity check. This will be done through the LDA (Latent Dirichlet allocation)
statistical model alongside cosine similarity. A corpus of nearly 30,000 movies scraped from
Wikipedia was used to train, evaluate and test the model. A basic GUI was also created to
showcase the main functions of the system as well as it give an idea how the end user will
interact with it.
Results : Through testing and evaluation it was revealed that the with big plots the results given
out by this system were either on par or in some cases offered better quality recommendations
than mainstream recommenders. However with smaller plots the results were inconsistent.
Discussion : This leads to the conclusion that a plot included recommendation system is not only
viable but with the added bonus of it being customizable leads the user to have an overall better
experience as they have more control over the recommendation process.
"
