面向不确定性数据的粒度模糊规则建模与优化方法研究

Massive data resources are rapidly accumulating in different fields, and it is urgent to understand the accumulated data by means of intelligent technologies such as data mining, knowledge discovery and reasoning. Traditional data mining techniques always process and analyze at the lowest level and the finest granularity of raw data, often consuming a lot of unnecessary computing resources and time. As to massive complex data (especially, the inaccurate, incomplete, inconsistent data), excessively fine data processing and data modeling is neither practical nor in line with human’ s cognitive habits, and it should have developed an effective solution. This study is focused on the uncertain-data-driven prediction and classification problems, and fully considers its characteristics. Inspired by the human’s cognitive mechanism, it builds a human-centric system based on granular computing theory and fuzzy theory for data mining purpose. Firstly, the uncertain data is abstracted into the description form of information granules, and the fuzzy rule-based inference system is constructed upon information granules. Then, by allocation of information granularity, the granular fuzzy rule-based model is constructed as long as the corresponding evaluation and optimization methods are designed. The quality of the system modeling and the space-time efficiency of the solution, as well as the fault tolerance, interpretability and robustness of the model reasoning are effectively improved.

在众多领域，海量的数据资源迅速积累, 迫切地需要借助数据挖掘、知识发现与推理等智能化技术理解积累的数据。传统的数据挖掘技术总是在最低层级和最细粒度的原始数据上进行处理分析，往往消耗大量不必要的计算资源和时间。过分精细的数据处理和数据建模，对于海量复杂数据，特别是不精确、不完备、不一致的不确定性数据的挖掘分析既不实际也不符合人们的认知习惯，亟需一种有效的应对方法。本项目面向不确定性数据驱动的预测与分类问题，充分考虑其特点，从人类的认知机制出发，以模拟人类层次模糊认知的粒计算理论和模糊理论为基础，建立一种以人为中心的数据挖掘感知方法。首先将海量数据抽象成信息粒的描述形式，并由信息粒支撑构建模糊规则推理系统，然后选择与问题相适应的粒度空间，通过设计相应的粒度配置策略，构建粒度模糊规则模型并设计相应的评估和优化方法，有效提高系统建模的质量和求解的时空效率，以及模型推理的容错性、可解释性和鲁棒性。

当前众多领域的海量数据资源迅速涌现, 迫切地需要借助数据挖掘、知识发现与推理等智能化技术理解数据。传统的数据挖掘技术总是在最低层级和最细粒度的原始数据上进行处理分析，然而过分精细的数据处理和数据建模，对于不精确、不完全、不一致的不确定性数据的挖掘分析还存在许多困难。本项目面向不确定性数据驱动的预测与分类问题，充分考虑其特点，从人类的认知机制出发，以模拟人类层次模糊认知的粒计算理论和模糊理论为基础，建立理论了一些的数据建模方法。针对不完全数据设计了信息粒填补和分类方法，以此为支撑构建模糊规则推理系统，并进一步提出了不确定性数据的信息粒表示与粒度模糊规则建模方法。通过分析模糊规则推理系统的构建过程，提出了模糊规则模型原型优化和粒度配置方法，以提高模型推理精度和鲁棒性。针对数据空间不一致，数据高维、有噪声和特征异构等特点，分别设计了基于协同模糊聚类的模糊规则建模方法、高维多变量数据驱动的粒度模糊规则建模方法、多视图数据的信息粒和分类建模方法。通过进行大量实验和典型案例分析，验证了上述方法能够有效提高系统建模的质量和求解的时空效率，以及模型推理的容错性、可解释性和鲁棒性。对于描述现实世界特别是不确定性环境下复杂的非线性系统具有十分重要的意义和先进性。