科技前沿识别体系中的机器学习应用问题研究

当前科技前沿识别研究方法难以得到更细粒度的分析结果，同时传统计量方法已不能够满足对当前来自网络的开源信息的情报挖掘需求，而机器学习方法可以实现数据细粒度的知识挖掘，因此成为解决科技前沿识别问题的重要手段。对2013—2021年中国知网和Web of Science(WoS)数据库收录的机器学习相关文献，在运用文献计量统计方法进行时间分布、研究主题及热点分析基础上，构建包含数据感知与处理层、情报计算和感知层、情报产品刻画层的开源情报环境下的科技前沿识别体系延伸架构，解读机器学习方法在各层次上的应用问题及关联关系，并提出不同层次需求发展的意见和建议；进而以7 944篇从WoS核心期刊库采集到的“深度学习”主题相关文献作为实验对象，主要针对数据处理中的知识单元构建进行论证。实证结果显示：从应用场景来看，多媒体信息处理的主题热度变化不大，智能机器人的主题热度逐年增高；从机器学习任务来看，目标检测和追踪的主题热度逐年降低，特征工程和数据分类则呈增长趋势。案例分析证明了所提出理论框架的科学性。     

Fair and Explainable Depression Detection in Social Media

Detection at an early stage is vital for the diagnosis of the majority of critical illnesses and is the same for identifying people suffering from depression. Nowadays, a number of researches have been done successfully to identify depressed persons based on their social media postings. However, an unexpected bias has been observed in these studies, which can be due to various factors like unequal data distribution. In this paper, the imbalance found in terms of participation in the various age groups and demographics is normalized using the one-shot decision approach. Further, we present an ensemble model combining SVM and KNN with the intrinsic explainability in conjunction with noisy label correction approaches, offering an innovative solution to the problem of distinguishing between depression symptoms and suicidal ideas. We achieved a final classification accuracy of 98.05%, with the proposed ensemble model ensuring that the data classification is not biased in any manner.