现代社会，社交媒体平台的广泛普及为用户的心理状况和心理需求带来了新的变化。关注用户在网络空间中的行为，对社交媒体用户进行心理计算，既有利于完善现有的心理学理论，也能更好地满足用户的个性化需求，提供主动关怀、改善用户体验，还有利于实现高效能、大规模、即时化的社交应用与社会管理。然而，当前的心理计算仍然存在挑战与不足，在数据集构建、特征与模型、测试与评价、实际应用这四个环节中分别面临着单一数据集的小样本标注问题、多模态数据的分析与建模问题、现实情况下的类别不平衡问题、心理计算与传统用户服务应用相结合的问题。本文针对以上四个环节和四个问题，系统性地研究了社交媒体用户的心理计算及其应用，取得了以下创新点：（1）领域自适应的跨平台心理计算。本文将领域自适应方法应用于心理计算，利用多个社交媒体平台的数据，通过统计性特征和跨语言预训练模型，探究不同平台的用户行为与心理计算的关联，发现领域特异和通用的用户特征，提出混合对齐的领域自适应网络，在特定平台的有标注训练数据不足的情况下，仍能取得良好的心理计算性能。（2）序列建模的多模态特征融合方法。社交媒体数据具有多模态、噪声多、内容隐含、上下文相关等特点，本文将数据驱动与知识驱动结合，定义多组手工特征和深度特征，引入注意力机制和序列模型，提出多模态序列网络，对社交媒体用户的时间线进行深入理解分析，充分挖掘和利用多模态数据进行用户的心理计算。（3）类别不平衡数据的心理计算。面向现实场景中的数据类别不平衡问题，本文构建了相应的大规模心理计算数据集，系统性地探究了不同程度的数据类别不平衡对心理计算性能的影响，证明了类别不平衡问题在心理计算中的重要性，并通过损失函数设计、输出阈值等方法，有效利用不平衡训练数据提升了现实场景下的算法性能。（4）以用户为中心的社交媒体应用。心理计算的最终目的是影响用户行为，提升用户体验。本文提出融合个性和情绪的注意力模型、多模态的注意力自编码器模型，利用社交媒体的多模态数据，发掘用户的个性和情绪等心理因素，实现以用户为中心的社交媒体应用，提升了应用的准确性，证明了心理计算在传统用户服务应用中的有效性。
Nowadays, the prevalence of social media platforms has made a difference to users‘ psychological states and needs. Psychological computing of social media users focuses on their online behaviors, and conduces to the development of existing psychological theories. It also helps to better satisfy users‘ personalized needs, provide pro-active care, improve user experience, and realize effective, large-scale and instant social applications and management. However, psychological computing still faces challenges, including insufficient labeled training samples of single dataset, analysis and modeling of multimodal data, class imbalance of data in real-world situations, and combination with traditional application services, which respectively involves the process of dataset construction, feature extraction and modeling, performance test and evaluation, and practical application. Aiming at these four problems and processes, this thesis systematically studies the psychological computing of social media users, and makes the following contributions:(1) Cross-platform psychological computing with domain adaptation. We apply domain adaption methods to psychological computing, and utilize data from multiple social media platforms. With statistical features and cross-lingual pretrained language models, we explore the correlation between users‘ behavior and mental states, discover universal and domain-specific features, and propose a Hybrid Alignment Domain Adaptation network (HADA), which achieves remarkable performance on certain platforms with insufficient labeled training samples.(2) Multimodal feature fusion with sequence modeling. Social media data is multimodal, noisy, connotative and contextual. We combine data-driven and knowledge-driven methods, define groups of handcrafted and deep features, employ attention mechanism and sequential models, and devise a Multimodal SequentialNetwork (MSN) to deeply understand and analyze users‘ timelines, and take full advantage of multimodal data for psychological computing.(3) Psychological computing on class-imbalanced data. Aiming at real-world situations with class-imbalanced data, we correspondingly construct a large-scale dataset and systematically explore the impact of data imbalance on psychological computing. We adequately reveal the significance of data imbalance issues, and propose a model with specialized loss function and output thresholding, which effectively improves the performance with imbalanced training data in real-world situations.(4) User-centric social media applications. The ultimate goal of psychological computing is to influence user behavior and improve user experience. We propose a Personality and Emotion Integrated Attentive model (PEIA) and an Attentive Multimodal Autoencoder approach (AMAE), which captures users‘ psychological factors (e.g. personality and emotion) with multimodal social media data. We realize user-centric social media applications with improved accuracy, and verify the effectiveness of psychological computing in traditional application services.
