الفهرس 
Chapter 1: IntroductionOnly 14 pages are availabe for public view
 |  | from | 128 |  |  |
| | | from | 128 | | |
Abstract
Currently, websites often overwhelm users with vast amounts of information, products, and services, and experts believe the overload of products can lead to incorrect decisions. Individuals want to quickly and easily find their target on the web, and recommender systems can limit the choices and recommend what best fits users’ needs. For example, Amazon recommends a wide range of products from various domains, such as mobile phones, computer accessories, electronics, books, and other items. There is a need for such e-commerce websites that cover a wide range of products and provide recommendations that take into account the relative importance of items in the domain, which we refer to as item’s utility. Making recommendations based on calculating the utility of items to the user is a promising research direction for developing e-commerce systems.
Developing recommender systems using utility mining can help with data sparsity and cold start issues. Furthermore, they enable customers to quickly discover the items with the greatest utility as a potential next purchase without wasting time with so many related searches, and they enable retailers to increase their profits. Since of their growing importance for both customers and retailers, these systems have caught the interest of both industry and academia. Industries are putting these systems to use in novel ways, while academics are investigating ways to improve them further. On the other hand, implementing recommender systems using utility mining comes with its own set of challenges. Firstly, memory and time constraints cause scalability issue when exploring the search space in a large database, especially since the downward closure property does not apply to the utility measure. Secondly, the majority of the patterns discovered by high utility pattern mining algorithms are poorly correlated or unreliable. These patterns are misleading and should not be used to make decisions. Thirdly, the observed progress in these systems is limited due to the issue of personalization limiting accuracy.
This thesis aims to improve the effectiveness and efficiency of recommender systems developed using utility mining. Firstly, we propose an effective approach called reliable utility-based pattern mining to mine itemsets with potentially high utility value in the future. An approach is valuable for decision-makers because it analyses consumption behaviors in buying the products and predicts interesting products in the future. Secondly, we introduce an efficient item decomposition clustering technique to avoid exploring the search space of all possible patterns. Additionally, providing a set of pruning strategies for the early elimination of unpromising patterns to reduce search space and time consumption. Thirdly, we propose an effective multi-objective recommendation approach that uses a wide range of interestingness criteria to meet the personalized demands of users while avoiding irrational inferences that could lead to misleading results. An approach is essential for users because rational decisions are required in e-commerce systems when selecting a product to purchase due to the need for objectivity and decision rationale tracking. The proposed approaches can improve the usefulness and applicability scale of utility-based recommender systems.
Experiments in synthetic and real-world databases show that the discovered patterns from the proposed reliable utility-based pattern mining approach are up to 47.60% more reproducible than state-of-the-art high utility pattern mining algorithms. Moreover, the proposed pruning strategies and item decomposition clustering technique can discard unpromising patterns from the search space, saving at least 98.2% of the search space. In terms of recommendation accuracy, the experiments demonstrate the importance of adapting objectivity and rationality in the recommendation process to improve the quality of recommendations while meeting the needs of as many users as possible. Furthermore, combining the multi-criteria weights of items obtained from the proposed multi-objective recommendation approach with the item-based collaborative filtering technique results in recommendations that are up to 60.95% more accurate than existing recommendation methods and improve user-level performance by up to 10.21% on dense datasets.