The research field of digital libraries involves a combination of different subfields from a wide variety of domains, and the usability and popularity of digital libraries depend on the satisfaction and user experiences of people while engaging with digital library systems. The digital library collections are nothing but large-scale resource search and retrieval systems, that are increasingly playing an important role in our knowledge society. They can be defined as catalogued and curated collections of different types of information resources, that are stored in a digital format in digital places. They can be viewed as online resource hubs, where it is possible to access information anytime and anywhere. It is important that every digital library needs to be equipped with a wide set of tools, assisting the users in their resource search and retrieval activities, and enhance their user-experience with these activities, leading to better usability of the system, and help users in finding relevant information and getting tailored information based on the personalized feedback, feelings and sentiments, ratings, preferences, and metadata build within the system. One of the tools that facilitates such a wide set of tailored and personalisation option, is called the recommender system, synonymously called the recommendation system, as well. The digital library systems equipped with recommendation system tools aim to provide personalisation features and assist users in finding relevant information based on their preferences, based on their previous choices, maintained in their respective profiles, and other similar users with either similar profiles or usage behaviour and pattern. The success of the recommendation process and tailored retrieval performance of the recommender system depends on the quality and quantity of the information used about the items and users, and their profiles stored from each interaction since the profiles represent the information needs of the users from the user-item interactions. To identify the needs of an individual user, and provide better personalisation, the design of the recommender system needs to be based on models that maintain detailed item descriptions and accurate users’ profiles, their preferences, and details of item-user interactions, and the performance and robustness of the entire recommendation system rely on the modelling accuracy on these factors. Due to massive information available about items and resources, often users are faced with ambiguity and diversity of information, due to the enormous growth of the Internet, with increasingly complex and massive volumes of the information, leading to difficulties in isolating the content that fits their needs. Due to this, users are not always certain of their information needs, nor do they know exactly how to describe what they want. The most challenging task involved in building tailored and personalized recommendations is acquiring information on user needs and their preferences, when there is limited information about users, particularly when the users are new, or do not engage with the system often. These problems make it difficult to profile users accurately and provide quality recommendations. Traditional approaches to address this issue in the research literature has been based on the concept of building taxonomical representation of the item and model the interaction information about each user and the item, and the relevance of each item to users and several other users of similar type. Such item representations or a taxonomical model can assist in determining users’ preferences. These models involve a hierarchical structure, comprising the coarse-grained representation to the fine-grained representations, for symbolizing the set of categories or topics, items, and users. However, due to the complex relationship between concepts, items, and users, from massive data sources, these taxonomic representation models are limited in their capabilities, and identification of each user’s information need is still a challenging task, particularly, when the user interaction with the items is limited, or the user happens to be new to the system. This thesis attempts to address some of these challenges, and proposes a novel computational framework based on deep machine learning techniques for building recommendation systems for digital libraries (modelled similar to a large-scale resource search and retrieval system), by including user feedbacks, preferences, explicit ratings, sentiments, and metadata information. In particular, the two deep learning models proposed in this thesis, the Multilayer Convolutional Neural Networks (MCNN) model and Collaborative Filtering based Multistage Deep Neural Network architecture (CFMDNN) model, allow better interaction and collaboration events to be captured between users’ feedbacks, explicit ratings, sentiments, and metadata information. The evaluation of the proposed computational framework based on several publicly available datasets shows significant enhancement in the recommendation systems performance, outperforming the related baselines. Further, the extension of these novel deep learning models for a new multilingual context (Arabic /English) digital library systems context, has led to promising results. In summary, this thesis makes significant contributions to recommender systems for digital library systems applications, with models based on multiple components including, explicit ratings, sentiment analysis based on text feedback, and metadata information for both Arabic and English languages contexts.
|Date of Award||2021|
|Supervisor||Girija Chetty (Supervisor) & Dat Tran (Supervisor)|