Topics in machine learning and data privacy

我的研究领域是机器学习/数据挖掘，也称为知识发现。我感兴趣的是增加现有学习系统的容量，并使它们更适合使用文本数据。我也在努力以负责任的方式使用这项技术：因此我对数据隐私感兴趣。 计算机使用个人数据来了解有关他们的新信息的能力是他们隐私的重要因素。我在这里的主要想法是开发工具，让人们控制他们的数据：每个人都应该能够决定谁可以使用他们的信息，以及用于什么目的。实现这一点的一种方法是控制应用软件，并允许它仅根据拥有信息的个人的授权访问和处理数据。我建议使用形式化的方法，即证明程序属性，来保证声称尊重用户授权的软件确实如此。我提出的研究的第二个主题是新的方法，应该表示文件，使他们可以有效地分类的计算机。我的方法是通过研究文档的结构、术语和它们所包含的术语的含义来进一步探索文档的语言属性。我特别希望这将提高专门分类任务的学习性能，特别是当数据稀疏时，例如在句子分类中。我感兴趣的最后一个领域是关系学习系统的可扩展性：使它们能够从大型的真实数据集中学习。为了科普信息溢出现象，我们必须采用知识信息提取和发现工具。在知识发现研究中的一个确定的差距是工具的可扩展性，特别是关系技术。然而，关系型数据在与社会最相关的应用中比比皆是：系统生物学和药理学，以及社交网络。因此，关系放大研究对知识发现的重要性。

My research is in the area of machine learning/ data mining, known also as knowledge discovery. I am interested in increasing the capacity of the existing learning systems, and in making them more suitable for using text data. I am also working on responsible ways of using this technology: hence my interest in data privacy.  Computer's ability of using data about individuals to learn new information about them is an important factor in their privacy. My main idea here is to develop tools that will give people control over their data: everyone should be able to decide who can use their information and for what purpose. One way to achieve this is to control the applications software, and allow it to access and process the data only according to the authorizations of individuals who own the information. I propose to use formal methods, i.e. proving program properties, to guarantee that software that claims to respect user authorizations, indeed does so. The second topic of my proposed research are new ways in which documents should be represented so that they can be efficiently categorized by a computer. My approach is to further explore linguistic properties of documents by looking at their structure, terminology, and the meaning of the terms they contain. I am particularly hopeful that this will improve learning performance of specialized categorization tasks, particularly when data is sparse, e.g. in sentence classification. My final area of interest is scalability of relational learning systems: enabling them to learn from large, real-life datasets. To cope with the information overflow phenomenon, we must employ knowledge information extraction and discovery tools. An identified gap in the knowledge discovery research is the scalability of tools, especially the relational techniques. And yet relational data abounds in the socially most relevant applications: systems biology and pharmacology, and social networks. Hence the importance of relational scaling-up research for knowledge discovery.