Bolstering pattern mining by taking into account relationships and uncertainty in data

总的来说，存储在数据库中的数据被认为是发生的事实和观察。这些记录通常也被认为是相互独立的。例如，医院数据库包含关于彼此独立的患者的信息，并且记录可以陈述在给定日期针对给定治疗所给予的特定诊断的访问。这些都是观察到的事实。数据挖掘和分析任务假设记录之间具有这种独立性，并认为记录的事实是确定的。然而，在许多真实的应用中，观测之间存在一些关系，并且这些关系是已知的，因此可以被记录。一个人是另一个人的朋友或家庭成员的事实，或者一个电话号码呼叫另一个电话号码的事实在记录之间创建了关系。这些是信息网络。在数据分析中忽略这些关系是错失了更好地了解数据的机会。此外，数据中存在不确定性，可能来自传感器等测量仪器或信息源所附的任何置信水平。例如，由传感器测量的温度可能不是确定的，而是假设在最小值和最大值之间的范围内;在十字路口处观察到的汽车的身份可能仅以一定的置信度来确定。这些是概率数据，因为属性的值可以附加概率水平。不确定性也可以归因于数据中的关系，以形成概率信息网络。不幸的是，大多数现有的数据挖掘方法假设数据的独立性和确定性。在有效地分析概率数据库或概率信息网络以在考虑到不确定性和关系的这种数据集合中发现有用的新知识或模式方面，已经做了很少的工作。 本提案的目的是设计有效和高效的技术，从概率数据库和概率信息网络中挖掘和学习，并显示这些技术在不确定性与数据收集密切相关的真实的领域应用中的使用和相关性，例如从文本数据中自动提取信息。

By and large, data stored in databases are considered facts and observations that took place. These records are also typically considered mutually independent. For instance, a hospital database contains information about patients that are independent from each other and a record could state a visit at a given date for a particular diagnostic for which a given treatment is given. These are all observed facts. Data mining and analysis tasks assume this independence between records and consider the recorded facts to be certain. However, in many real applications some relationships exist between observations and are known and thus could be recorded. The fact that a person is a friend of another person, or a family member, or the fact that a phone number calls another telephone number creates a relationship between records. These are information networks. Ignoring these relationships in data analysis is a missed opportunity to get better insights about the data. Also, uncertainties in data exist and can come from the measurement instruments such as sensors or any confidence level attached to the source of information. For instance, the temperature measured by a sensor may not be certain but assumed in a range between a minimum and a maximum; the identity of a car observed at an intersection may only be ascertained with some level of confidence. These are probabilistic data because the value of an attribute could be affixed with a probability level. Uncertainty can also be ascribed to relationships in data to form probabilistic information networks. Unfortunately, most existing data mining approaches assume independence and certainty of data. Very little work has been done on effectively analyzing probabilistic databases or probabilistic information networks to discover useful new knowledge or patterns in such data collections that takes into account the uncertainties and the relationships. The purpose of this proposal is to work on devising effective and efficient techniques to mine and learn from probabilistic databases and probabilistic information networks, and show the use and relevance of these techniques in real domains applications where uncertainty is germane to the data collection, such as in automated information extraction from textual data.