| Summary: | To alleviate the data sparsity and cold start issues in recommendation, many researchers leverage user relationships (e.g. social network) to improve the accuracy of recommender systems. However, social connections may not be available in many real systems, whereas item relationships are much easier to obtain but lack of study. Therefore, in this disser- tation, we focus on exploiting item relationships for effective recommendation.
In real applications (e.g. Amazon), items are often organized by category, which is a popular way to define item relationships. Based on the assumption that users tend to have similar preferences towards items that belong to the same category, plenty of category-aware recommendation methods have been proposed. But, they mainly consider categories that are organized in a flat structure, where categories are independent and in a same level. In fact, categories can be also organized in a richer knowledge structure, i.e., category hierarchy (CH), to describe the inherent correlations among different categories, which might be more helpful to enhance the recommendation performance.
In order to take advantage of CH for better recommendation, we first propose a novel matrix factorisation framework with recursive regularisation – ReMF. It not only jointly models and learns the influence of organised categories on user-item interactions, but also provides characterisation of how different categories in the hierarchy co-influence the modelling of user-item interactions. Empirical results show that ReMF consistently outperforms state-of-the-art category-aware recommendation methods.
Despite the success of ReMF, we notice that all CH based methods merely focus on the influence of vertically affiliated categories (i.e. child-parent) on user-item inter- actions. The relations of horizontally organised categories (i.e. siblings and cousins) in CH, however, have only been little studied. We show in real-world datasets that category relations in horizontal dimension can help explain and further model user-item interactions. To fully exploit CH, we further devise a unified matrix factorisation framework – HieVH, that seamlessly fuses both vertical and horizontal dimensions for effective recommendation. Extensive validation on real-world datasets demonstrates the superiority of HieVH against state-of-the-art algorithms. An additional benefit of HieVH is to provide better interpretations of the generated recommendations.
Recently, representation learning (RL) has proven to be more effective than matrix factorisation in capturing local item relationships by modelling item co-occurrence in individual user’s interaction record. We further design a unified multi-level RL based Bayesian framework – MRLR, thus benefiting from RL. By fusing item category, MRLR captures fine-grained item relationships from a multi-level item organization: items in local context (i.e., item co-occurrence relations), items affiliated to the same category, and items in user-specific ranked list. To the best of our knowledge, we are the first to investigate item category from the perspective of multi-level RL. Experimental results on multiple datasets show that MRLR consistently outperforms state-of-the-art algorithms.
Besides, with the development of semantic web, the knowledge graph (KG) has re- cently attracted a considerable amount of interest in recommendation, as it connects various types of features related to items (e.g., the genre, director, actor of a movie), in a unified global representation space. Utilising such kind of heterogeneous connected information facilitates the inference of subtler item relationships from different perspectives, which are difficult to uncover with the homogeneous information (e.g., item category) only. To fully exploit the heterogeneous information encoded in KG for better recommendation, we propose a KG embedding framework – RKGE based on a novel recurrent network architecture that automatically learns semantic representations of entities and paths. In particular, RKGE learns the semantic representations of entities and paths between them via a batch of recurrent networks, and seamlessly integrates them into recommendation. Furthermore, it employs a pooling operator to discriminate the saliency of different paths in characterising user preferences over items. Empirical study demonstrates that RKGE outperforms state-of-the-art algorithms. In addition, we show that RKGE provides meaningful explanations for the recommendation results.
To sum up, in this dissertation, we propose a series of recommendation approaches by exploiting auxiliary item relationships to deal with the data sparsity and cold start problems of recommender systems, which are natural but novel extensions of existing proposals for effective recommendation.
|
|---|