在信息爆炸时代，用户常常难以从海量互联网数据中找到感兴趣的内容，而个性化推荐系统则能够根据用户建模结果，主动发掘出用户可能感兴趣的信息并将其推荐给用户，在满足用户信息需求的同时提升用户体验。因此，推荐系统成为了互联网中一类不可或缺的服务。推荐系统的研究往往离不开用户建模，因为只有更有效地开展用户建模，系统才能更准确地理解用户兴趣，为其提供精准服务。用户建模与推荐系统的构建依赖于合理地使用用户的已有信息。而用户往往会使用种类繁多的网络服务和应用，然后在这些异质环境产生各类跨模态、跨领域或跨平台的异质用户信息。这些信息来自不同的数据场景，分布和特性往往有差别。因此，如何整合多源异质数据已成为用户建模与推荐方法研究中的重要挑战，但也为研究带来了新的机遇：首先，现有的用户建模研究工作主要利用文本模态信息，较少工作使用了多模态异质信息；其次，大部分推荐算法主要使用推荐场景中的用户信息，冷启动问题将制约算法的表现；第三，知识信息（如：商品的互补关系）可能能帮助改善推荐结果，但这些知识信息与用户信息为异质信息。针对以上挑战，本文开展了面向异质环境的用户建模与推荐研究：(1)我们针对用户属性建模，设计了多模态多粒度用户角色建模算法，并在真实数据集上识别正确率达到90.4%。通过深入分析用户角色，我们还结合心理学理论改进了已有的用户建模与应用算法；(2)我们尝试精准建模用户主观偏好，提出了结合跨平台信息的推荐方法。本工作将用户的跨平台信息应用到兴趣建模，提出的算法在豆瓣数据集上较基准算法均方根误差降低了4.2%，并能对冷启动用户进行准确预测；(3)我们致力改进用户需求预测算法，设计了引入知识图谱推理的新算法。本工作基于知识图谱挖掘了用户交互历史与待推荐内容的关系，进而刻画用户在当前场景的信息需求并完成推荐。算法在亚马逊数据集上召回率平均提升了4.39%。本论文关注于融合异质信息来改进用户建模与推荐方法，主要创新点如下：（1）首次设计了融合多模态信息的多粒度角色识别框架，并将用户角色特征用于观点和偏好预测；（2）提出的跨平台跨领域异质信息推荐方法能够有效应对冷启动问题，发现“推荐领域无关”的信息同样有助于兴趣建模；（3）设计了结合知识图谱中知识信息的推荐方法，通过推理规则赋予了推荐方法更强的可解释性。这些方法具有良好的鲁棒性和可扩展性，能够改善真实场景下推荐系统的表现。

In the era of information explosion, personalized recommender systems have become indispensable online services. Based on the results of user preference modeling, recommender systems will extract information that may be of interest to every user from the massive information and recommend it to the user. Better user modeling will result in a better understanding of user interests, and then recommender systems can provide users with better recommendation results to meet their demands based on their interests.Since users often use multiple services and applications on the Internet, they will generate different types of heterogeneous information: multi-modality, cross-domain or cross-platform information. As heterogeneous information is derived from heterogeneous environments, the distributions and characteristics of it are often different. Only by comprehensively using this information can achieve better users modeling and recommendation results. Thus, how to integrate these multi-source heterogeneous data has become a big challenge in the research of user modeling and recommender systems, but it also brings new opportunities: Firstly, existing user modeling studies usually based on only single modality data, especially textual data. While other modality information generated by users are often ignored. Secondly, most previous recommendation methods focus on making use of existing features on the recommender systems, which often suffer from the cold-start problem. Thirdly, knowledge graph, which contains a large amount of knowledge information, can be helpful for the recommendation. While the structured knowledge information and user behavior information is heterogeneous information.In order to overcome these challenges, this paper conduct research on user modeling and recommendation under heterogeneous environments: (1) We start our research from modeling the objective attributes of users, the proposed multi-modality multi-granularity algorithm achieves 90.4% role recognition accuracy on real datasets. Besides, inspired by psychology theories, we modify some algorithms to incorporate user role features and achieve good performance in experiments. (2) We try to model users' subjective preference more accurately with cross-platform user features. A new framework which is able to incorporate cross-platform heterogeneous information is proposed. The designed method is flexible to work with existing rating prediction algorithms. The enhanced algorithms get significant improvements in Douban datasets, the root mean square error is reduced by 4.2%. Besides, our methods are able to cope with cold start users. (3) We are committed to integrating knowledge graph and reasoning with recommender systems. We combine the heterogeneous information from knowledge graph to mine the features between items, and then the association feature enhanced recommendation algorithms outperform the original methods significantly, the averaged recall@5 improvement is 4.39% on Amazon dataset.We focus on how to effectively integrate heterogeneous data to improve user modeling and recommendation performance in this study. The main contributions are as follows: (1) A multi-granularity role recognition framework with multi-modality information is designed for the first time, and user role features are introduced for user preference and attitude prediction; (2) The proposed cross-platform cross-domain heterogeneous recommendation algorithm is able to deal with the cold-start problem, and our study verify that ``off-topic" information is also helpful for preference modeling; (3) The proposed framework that combines knowledge information is able to give an explanation of the recommended results. These proposed methods have good robustness and scalability, and can effectively improve the performance of recommendation systems in real-scenarios.