工作记忆理论驱动的社交媒体用户画像自然属性预测模型

Working Memory Theory Driven Natural Attribute Prediction Model for Social Media User Profiling

利用社交媒体用户生成的内容构建用户画像系统,为个性化服务和电商平台的精准营销提供支持,这是社交媒体分析领域重要的研究方向之一.文中研究用户以时间线方式发布内容形成篇章级长度的多模态数据,并针对由此给用户画像自动构建带来的挑战展开分析.以用户性别和出生年代为主的自然属性为目标,研究如何有效处理与分析社交媒体用户发布的篇章级多模态数据,提出基于Moments的社交媒体用户画像自然属性预测模型.在认知心理学启发下,通过工作记忆理论设计数据分块方式,缓解以拼接和滑动窗口为主的传统方法存在的语意破坏和语篇合成的问题.再针对用户在发表社交内容时存在的模态偏好问题,综合注意力平衡模态内数据和模态间数据对预测任务贡献度的差异,实现用户画像中自然维度属性的预测.实验表明,文中模型在用户性别和出生年代预测任务上的准确性较高.

Constructing user profiling systems using contents generated by social media user can offer personalized services and precise marketing for e-commerce platform.It is a significant research direction in the field of social media analysis.In this paper,the document-level multimodal data formed by users publishing content chronologically is studied,and the challenges brought by that to user profiling are analyzed.Aiming at the natural attribute primarily related to user gender and birth year,how to deal with and analyze the document-level multimodal data posted by social media users efficiently is studied as well.A natural attribute prediction model for social media user profiling is proposed.Inspired by cognitive psychology,an effective data chunking method is designed via working memory theory to alleviate the problems of semantics broken and synthetic discourse in traditional methods.To solve the problem of user content preference,an attention mechanism is employed to balance task contributions between intra-modal and inter-modal data.Experiments show that the proposed model is superior in user gender and birth year prediction.