Recommendation System for Next Generation of Smart TV
MetadataShow full item record
This research thesis is dedicated to the design of recommendation system for television. Nearly every household has at least one TV and pays a broadcast provider for a number of channels.. Orientation in program offer can be challenging and having a real time overview of hundreds of channels is impossible. This means the money paid for service is ineffective if not wasted completely. Recommendation systems are designed to save time and effort when searching for a suitable content. The system learns user preferences from past observations and suggests a content fitting these preferences. There are 5 basic recommendation techniques, each using different kind of knowledge for their prediction and their hybrid combinations. In this thesis, two most commonly implemented approaches are described in detail, namely, content and collaborative filtering. The main focus of this work is on handling categorical data as electronic program guide can provide a rich description of a program. State-of-the-art methods associated with this two recommendation approaches are described and some of them are extended to improve their performance. For my design, I choose to apply a probabilistic approach allowing comprehensive manipulation of the categorical data as well as providing insight into feature relationships of the content description. Graphical models meet all the requirements and because of this reason they become the primary approach the design on is built on. A novel approach based on transfer learning is applied to the graphical network in this thesis. This approach is able to benefit from user group information, therefore overcoming the issue of insufficient user data. The proposed recommendation system is applied to a television environment involving the emerging technique of hybrid broadband and broadcast (HBB) transmission of TV content. HBB is a standardized platform combining and harmonizing streams from broadband and broadcast sources allowing a simple implementation of entertainment services to enhance the user experience. Recommendation engine is one of the interactive services in this framework allowing the user to have an overview of favourite programs. The recommendation is made based on the estimation of a rating prediction. The item with the highest predicted rating is then recommended. This makes an accurate rating prediction crucial for the performance evaluation of the model. Because of this, beside the commonly used mean absolute error (MAE), a new metric to measure the performance of a recommendation engine is proposed which focuses on the importance of rating prediction. Experiments are performed on real world data set provided by Yahoo Labs. It is a collection of movies with their description categorized in a number of features and user ratings. The item description is often incomplete with many feature values missing. This is common for many data sets. Another typical issue encountered by this data set is the sparseness of the user-item matrix and item-feature matrix. It is beneficial if the recommendation system is designed in a way that these issues are either minimized or the model is robust enough that the system design is not affected by them. The model proposed in this work incorporates a method for missing value estimation and does not suffer from the sparsity issue.
Thesis (PhD Doctorate)
Doctor of Philosophy (PhD)
School of Info & Comm Tech
The author owns the copyright in this thesis, unless stated otherwise.