Abstract/Summary

Recommender systems are crucial for addressing information overload by providing users of online platforms, applications, or services with relevant item suggestions. Visually-aware systems enhance recommendations using images in various domains, e.g., fashion and entertainment. However, most visually-aware recommender systems lack explainability due to their reliance on black-box approaches such as deep-learning techniques. This thesis addresses the need for accurate and explainable visually-aware recommender systems by integrating visual information into Heterogeneous Information Networks (HINs). Three main research questions arise: How can the effective integration of visual information into HINs be achieved? How can the development of explainable visually-aware recommender systems based on HINs be approached? What methods can be used to enhance the explainability of HIN-based recommender systems? To achieve these, this thesis proposes a three-part solution. Firstly, visually-augmented HINs are constructed by introducing visual factor nodes and relations, generated from various image features. A user representation learning method is introduced to combine semantic and visual preferences. Secondly, a Scalable and Explainable Visually-aware Recommender System (SEV-RS) framework is presented. SEV-RS utilizes meta-paths for explainability, enhancing recommendations using scalable feature extraction from visually-augmented HINs. Lastly, a meta-path translation model is introduced to improve the explainability of meta-path based systems, enhancing the performance of the proposed framework. Extensive experiments with real-world datasets in movies and clothing domains assessed the effectiveness of visually-augmented HINs. These augmented HINs, along with the proposed user representation learning method, were employed in a recommendation model using Collaborative Filtering with K-Nearest Neighbors (CF-KNN). This model was compared with other CF-KNN models based on regular HINs. The findings indicate the efficacy of visually-augmented HINs and the representation learning approach in enhancing CF-KNN. They also exhibit the practicality of using visually-augmented HINs in state-of-the-art recommender systems. The evaluation of SEV-RS involved comparisons with state-of-the-art models using real-world and synthetic datasets. The results reveal that SEV-RS yields recommendations with high accuracy and explainability while requiring notably less computational time compared to other deep-learning models. The proposed meta-path translation approach was evaluated on two newly generated datasets derived from real-world recommendation data. It was compared with various sequence-to-sequence models in the task of translating a given meta-path to a group of meta-paths with higher explainability. The experiments validate its capability to generate more comprehensible alternative explanations for complex meta-paths. These approaches create pathways for developing recommender systems that address users’ visual preferences and offer explanations based on understandable meta-paths. Implementing these methods could enhance user experiences, leading to more engaging and personalized recommender systems, with potential adaptability for improving other graph-utilizing explainable artificial intelligence applications.