Adversarial knowledge discovery

对抗性知识发现从数据中建立模型，在这种情况下，进行建模的人的利益与一些被建模的人的利益不一致。这包括执法、反恐、边境管制、欺诈、反洗钱，以及越来越多的主流领域，如客户关系管理。犯罪活动每年给加拿大造成的损失可能高达GDP的6%，数十亿美元；一些犯罪和恐怖主义还会造成财产损失、伤害和生命损失。 目前，从对手在数据中留下的痕迹来理解他们的行动的过程是由分析师推动的，他们询问数据，寻找他们认为可能重要的模式。这是一项繁重的工作，需要分析人员具备技能、经验和创造力。通过增加归纳成分，他们的结果可以得到显著改善--根据数据以算法构建可信的模型，并要求分析师对其进行评估。 这在主流知识发现中很常见，但在对抗性环境中更难发现，因为可能的模式更难发现：因为对手正在积极地尝试改变他们的活动，因为他们试图隐藏自己，因为他们可能试图操纵分析过程。我在这一领域的下一阶段工作将解决四个问题：(1)找到算法来对数据记录进行排名，而不需要假设，基于“兴趣度”，以便分析师的注意力可以集中在关键的少数几个方面；(2)从文本中更深层次地提取信息，考虑我们作为人类是如何生成语言并使其成为算法的；(3)从图形结构的数据中提取更丰富的信息(例如，基于人与人之间的互动的关系联系；以及(4)通过算法发现数据中簇和其他结构的潜在含义。 在解决这些问题方面取得进展将使加拿大更加安全，将减少我国人民遭受犯罪和恐怖主义暴力的风险，并降低犯罪和欺诈的重大经济成本。

Adversarial knowledge discovery builds models from data in settings where the interests of those doing the modelling are not aligned with the interests of some of those being modelled. This includes law enforcement, counterterrorism, border control, fraud, anti-money-laundering and, increasingly, mainstream domains such as customer relationship management. Criminal activity costs Canada perhaps as much as 6% of GDP, billions of dollars per year; and some crime and terrorism also has costs in property damage, injuries, and lives lost. At present, the process of understanding the actions of adversaries from the traces they leave in data is driven by analysts who interrogate the data for patterns that they think may be significant. This is laborious and requires analysts to be skilled, experienced, and creative. Their results can be improved substantially by adding an inductive component -- constructing plausible models algorithmically from the data, and asking analysts to assess them. This is commonplace in mainstream knowledge discovery, but more difficult in adversarial settings because likely patterns are harder to discover: because adversaries are actively trying to vary their activities, because they are trying to conceal themselves from analysis, and because they may be trying to manipulate the analysis process. The next stage of my work in this area tackles four problems: (1) finding algorithms to rank data records without assumptions, based on 'interestingness' so that analyst attention can be focused on the critical few; (2) extracting information from text in a deeper way,taking ideas about how we, as humans, generate language and making them algorithmic; (3) extracting richer information from graph-structured data (for example, relational connections among people based on their interactions; and (4) algorithmically discovering potential meanings of clusters and other structures in data. Progress in solving these problems will make Canada safer, will reduce the risks to our population from criminal and terrorist violence, and reduce the substantial economic cost of crime and fraud.