景点推荐系统通过分析用户的行为、兴趣和需求，提供个性化的景点推荐结果，可以有效缓解海量旅游资讯给用户带来的信息过载问题。相比于电商、电影推荐等，旅游领域的交互数据较少，数据稀疏性问题严重；同时，旅行者需要更加多样化的推荐结果，以及更加具体的推荐解释。本文聚焦于旅游景点推荐领域，针对景点推荐系统所存在的稀疏性、多样性、解释性等需求与问题进行探索，本研究主要工作与成果如下： 1）针对旅游领域的数据稀疏性问题，本文提出了一种新的旅游知识图谱构建范式——旅游事件知识图谱。不同于通用知识图谱侧重于描述景点的静态知识，旅游事件知识图谱侧重于描述旅游事件中游客的行为与体验。本文构建了真实的景点评分数据集，并以上海市为例构造了景点通用知识图谱和旅游事件知识图谱样例以应用于推荐模型中。实验结果证明旅游事件知识图谱应用于推荐模型中可以为推荐准确率带来平均2%的提升，优于景点通用知识图谱的效果。 2）针对景点推荐的多样性需求，本文提出了一种多粒度兴趣推荐模型，将对用户历史行为的“记忆”与对用户兴趣的“泛化”相结合。模型中的细粒度兴趣模块用于保留对用户历史行为的真实记忆；粗粒度兴趣模块则用于抽象出多个更加泛化的高级兴趣。两个模块从不同的粒度上相互补充，平衡推荐结果的准确性与多样化。采用五个真实世界的推荐数据集开展了充分的实验，证明多粒度兴趣推荐模型在准确性指标上相比最先进的点击率预测模型提升了1-2%；在多样性指标上的提升则超过了10%。 3）针对用户对景点推荐结果的解释性需求，本文在前两个工作的基础上提出了一种知识增强的多兴趣学习方法，将知识图谱中丰富的细粒度兴趣属性融入到多兴趣推荐模型中，以“属性”作为多兴趣学习的最小粒度，使得对用户兴趣的聚类过程更加精细，对兴趣向量所代表的用户兴趣、候选物品所包含的兴趣也可以提供更加具体的解释。实验结果表明知识增强的多兴趣学习方法可以进一步将多样性指标提升7.7%，本文也通过案例研究阐述了该方法在兴趣学习的过程中具有良好的可解释性。 综上所述，本文运用深度学习相关知识来解决旅游领域中景点推荐系统的需求与问题，针对该任务的特点提出了相应的方法，改进推荐效果。
Base on users’ behaviors, the attraction recommendation system provides personalized recommendation services, alleviating the information overload problem brought to users by the huge amount of travel information. Compared with the fields of movies and music, the tourism field has serious data sparsity characteristics; at the same time, travelers need more diverse recommendation results and more specific recommendation explanations. This paper focuses on the field of tourist attraction recommendation, and explores the needs and problems of sparsity, diversity, and explanation for the attraction recommendation system, and accomplishes the following work: 1) To address the data sparsity problem in the tourism field, this paper proposes a new paradigm for building tourism knowledge graphs--Tourism Event Knowledge Graph(TEKG). Different from the generic knowledge graph which focuses on describing the static knowledge of attractions, TEKG focuses on describing the behaviors and experiences of tourists in tourism events. This paper constructs a real-world attraction rating dataset, and takes Shanghai city as an example to construct samples of the generic knowledge graph and TEKG. The experimental results demonstrate that TEKG can improve the recommendation accuracy by 2% on average when applied to the recommendation model. 2) For the diverse needs of attraction recommendation, this paper proposes a multi-granularity interest recommendation model, which combines the “memory” of users’ behaviors with the “generalization” of users’ interests. The fine-grained interest module in the model is used to retain the real memory of the user’s historical behavior; the coarse-grained interest module is used to abstract multiple more generalized and high-level interests. The two modules complement each other at different levels of granularity to achieve a balance between accuracy and diversity of recommendations. Five real-world recommendation datasets are used to conduct sufficient experiments, which demonstrate that the multi-granularity interest recommendation model improves 1-2% in accuracy metrics; and improves more than 10% in diversity metrics. 3) Based on the previous two works, this paper proposes a knowledge-enhanced multi-interest learning method, which integrates the rich fine-grained interest attributes in the knowledge graph into the multi-interest recommendation model, and uses “attributes” as the minimum granularity of multi-interest learning, so that the clustering process of user interests is more refined. The process of clustering user interests is more refined, and more specific explanations can be provided. The experimental results prove that the knowledge-enhanced multi-interest learning method can further improve the diversity metrics by 7.7%, and this paper also illustrates the good interpretability of the method in the process of interest learning through case studies. In summary, this paper uses deep learning knowledge to solve the problems existing in the attraction recommendation system, and proposes corresponding methods to improve the recommendation effect.
