Formal concept analysis and text mining

我们的研究项目涉及两个主要方向。形式概念分析（FCA）和文本挖掘。我们的第一个研究方向是在FCA和应用领域，特别是数据挖掘应用领域开展正在进行的工作。FCA可以看作是一种分层概念聚类方法，它揭示了基于伽罗瓦（或概念）格的形式框架的表格数据中的共性和关系的重要模式。在80年代早期由Wille和我自己开创，这个领域得到了很大的扩展，在过去的十年里，它已经成为一个建立良好的学科，具有详细的理论基础，大量的应用（数据挖掘，生物信息学，软件工程，信息检索，语言学，心理学，…）和软件工具（商业和开源），以及一个不断增长的国际社区。在下一个授权期内，我们计划关注以下主题中的数据挖掘应用：多层关联规则数据挖掘压缩表示、复杂数据挖掘、近似压缩表示和用于数据仓库性能优化的增量频繁模式挖掘。我们的第二个主要研究方向是文本挖掘和信息检索应用。许多可用的电子信息是以文本的形式存在的。网络就是一个明显的例子。文本挖掘涉及从文本数据中提取有用的表示。这些表示特别用于信息检索和文本分类应用程序。典型的单术语索引和加权方案无法系统地利用通常更接近底层概念的共现模式。我们正在研究几种提取重要的共现模式的方法。更具体地说，我们的工作利用遗传算法，人工神经网络和频繁集挖掘算法。

Our research program addresses two main directions. Formal Concept Analysis (FCA) and text mining. Our first research direction is to pursue ongoing work in the field of FCA and applications, in particular data mining applications. FCA can be viewed as a hierarchical conceptual clustering approach that reveals the significant patterns of commonalties and relationships within tabular data based on the formal framework of Galois (or concept) lattices. Pioneered by Wille and myself in the early eighties, this field has largely expanded and it has become over the last ten years a well established discipline with an elaborated theoretical basis, a large array of applications (data mining, bioinformatics, software engineering, information retrieval, linguistics, psychology, ...) and software tools (commercial and open source), and a growing international community. In the next grant period, we plan to focus on data mining applications within the following main themes: multilevel association rule data mining condensed representation, complex data mining, approximate condensed representations and incremental frequent pattern mining for performance optimization in data warehouses.Our second main research direction is text mining and information retrieval applications. Much of the available electronic information is in the form of text. The Web is an obvious case. Text mining is concerned with extracting useful representations from textual data. These representations are used in information retrieval and text classification applications in particular. Typical single term indexing and weighting schemes suffer from their inability to systematically exploit co-occurrence patterns that are often more close approximations of the underlying concepts. We are working with several approaches that extract significant co-occurrence patterns. More specifically, our work exploits genetic algorithms, artificial neural networks and frequent set mining algorithms.