Recommender systems have been developed at different levels with different approaches to help resolve the challenge of choice making due to the abundance and variety of information. The aim however remains the same — to provide accurate recommendations for users. Traditional approaches to recommender systems have mainly focused on collaborative filtering or content-based filtering. More recent approaches have explored hybrid models which combine the collaborative and content-based filtering methods. They however fall short in addressing the issues of personalization, cold start, diversity and explainability in recommender systems.This thesis intends to address these shortfalls by utilizing the advances in Natural Language Processing and Machine Learning. First, in order to personalize users' recommendations, it is essential to consider their personalized preferences on non-functional attributes. However, to increase recommendation accuracy, it is essential that recommendation systems include users' evolving preferences. Existing recommendation systems fail to thoroughly capture users' dynamic preferences for personalized recommendation. This work proposes a method to personalize users' recommendations based on their dynamic preferences on non-functional attributes.Secondly, we use side information to resolve cold start and data sparsity limitations in RS, in order to improve their accuracy. Knowledge graphs (KGs) have shown to be very valuable source of side information because they allow hybrid graph-based recommendation methods comprising both collaborative and content information. Using KGs as side information also helps us to find latent connections between the entities in a dataset, to improve recommendation precision and bring explainability in recommendations.Finally, with the use of transformer pre-trained models, we introduce the ELECTRA-KG (Efficiently Learning an Encoder that Classifies Token Replacements Accurately) which utilizes a custom KG built with the existing items in a data base and treats entities, relations and triples as textual sequences thereby turning KG completion into a sequence classification problem. It uses the description of the entities together with their relations in the KG to compute a scoring function of the triples. By doing so, we can determine the plausibility of a triple or a relation from our fine-tuned model. Experiments and evaluations on real world data shows the effectiveness and accuracy of our proposed models.
展开▼
机译：推荐系统在不同层面上以不同的方法开发，以帮助解决由于信息丰富和多样化而导致的选择挑战。然而，目标保持不变——为用户提供准确的建议。传统的推荐系统方法主要集中在协同过滤或基于内容的过滤上。最近的方法探索了混合模型，这些模型结合了协作和基于内容的过滤方法。然而，它们在解决推荐系统中的个性化、冷启动、多样性和可解释性问题方面存在不足。本论文旨在通过利用自然语言处理和机器学习的进步来解决这些不足。首先，为了个性化用户的推荐，必须考虑他们对非功能属性的个性化偏好。然而，为了提高推荐的准确性，推荐系统必须包括用户不断变化的偏好。现有的推荐系统无法全面捕捉用户的动态偏好，从而实现个性化推荐。这项工作提出了一种方法，根据用户对非功能属性的动态偏好来个性化用户的推荐。其次，利用侧面信息解决RS中的冷启动和数据稀疏性限制，以提高其准确性。知识图谱 （KG） 已被证明是非常有价值的辅助信息来源，因为它们允许基于混合图的推荐方法，包括协作和内容信息。使用KG作为辅助信息还有助于我们找到数据集中实体之间的潜在联系，以提高推荐精度并带来推荐的可解释性。最后，通过使用 transformer 预训练模型，我们引入了 ELECTRA-KG（Efficiently Learning an Encoder that Classizes Token Replacements），它利用数据库中现有项目构建的自定义 KG，并将实体、关系和三元组视为文本序列，从而将 KG 完成转化为序列分类问题。它使用实体的描述以及它们在 KG 中的关系来计算三元组的评分函数。通过这样做，我们可以从微调模型中确定三元组或关系的合理性。对真实世界数据的实验和评估表明了我们提出的模型的有效性和准确性。
展开▼
