Declarative context-aware recommendation

Recommender systems aim to assist users in identifying and retrieving particular items from a large collection of items that are relevant to their preferences. The preferences of users are usually learned from their previous transactions. User preferences are difficult to understand because they are influenced by a variety of obscure situations or contexts that exist when the user makes a decision. Contexts are dynamic, and each user's value of each context varies, resulting in a dynamism of user preferences. Context-aware recommender system (CARS) has been proved to increase the quality of recommendations by incorporating contextual information and personalising them for each user. Personalising contexts for each user and incorporating them into the process of recommendations are two main issues of CARS. In this thesis, we address these two issues in order to improve the performance of recommender system in terms of effectiveness and efficiency. Specifically, depending on the characteristics of the available data, we adopt three distinct approaches to tackle these issues. The first data characteristic is associated with the situation when users preferences are accessible through user reviews of purchased items. Generally, these reviews describe the user's rating of the item. In this research, we propose UW-NMF, a new topic modelling for identifying the users' contexts from their declared reviews. UW-NMF personalises and identifies the most relevant contexts for each user. Then, we use XGBoost to incorporate the personalised contexts into the recommendation generation process. We named the first model, D-CARS. We show that D-CARS can increase the quality of recommendation for the most active users through comprehensive studies. We also propose DecPro-CARS, which seamlessly incorporates not only the user profiles but also the item profiles into the recommendation model. DecPro-CARS aims to work on all users, whether they are active or not. DecPro-CARS then matches the user and item profiles to recommend the most relevant items to the user. In most cases, user preferences may be gleaned from their activity history. However, due to the dynamic nature of the contexts in decision making, a new preference may emerge that has not yet been captured, which is our second data feature. Given this characteristic, we propose a new method for capturing user instant preferences by leveraging user interactions with the system. Keeping in mind that each context has a varied priority for each user, we first provide a novel MF approach termed PW-CAMF to learn the importance of contexts for each user. We collect user instant preferences on the top of PW-CAMF by asking them to provide input to the system using one of three proposed question selection strategies: MaxWeight, MaxIUWF, or greedyIUWF. The user response is fed back into the model via context learning, and then the most relevant items are returned to the user. The third data characteristic occurs when the order of transactions influences the user's preferences. A decision is made in this case based on previous or future user actions. With this data characteristic in mind, we propose a Context Graph (CGNMPI) and a Transaction Context Graph (TCG) to capture user contexts sequentially. In particular, CGNMPI models the relationship between contexts, allowing us to mine all relevant contexts within a given context. In the meantime, TCG models the importance of contexts in each user transaction, capturing the situations in which a user transacts. We incorporate learned user preferences into the recommender system by using a bidirectional transformer called BERT. Declarative recommendationThis thesis examines the declarative contexts in which users conduct transactions, particularly in terms of user reviews and/or predefined implicit or explicit contexts. We propose three novel recommendation approaches that are based on the characteristics of the available data. The extensive experimental results show that the proposed approaches can address the objectives of this study and outperform the existing competitors.