Prescriptive Analytics over Graphs, Streams, and Sequences

机器学习(ML)的进步推动了预测分析和推荐系统的许多成功应用。要将这些成功推向下一个水平，需要规范分析，具有两个关键功能：i)对模型所做的预测进行查询；ii)建议干预行动，如果采取干预行动，可能会导致模型预测的预期结果。该计划的一个主要长期愿景是开发一个框架、理论、模型和算法来实现这些功能，具体体现在三个关键应用上：A)病毒式营销(VM)；B)错误信息遏制；以及C)对医疗轨迹数据库的干预建议。我们精心选择了应用程序，使它们既相关又多样，它们以图形、流和序列等异类数据为特征，并将用于说明我们将要开发的模型和技术的一般性。给定正在进行的VM活动的快照，营销人员会想知道结果是否有可能达到销售目标。此外，如果产品遇到来自竞争对手公司的新竞争，她可能想要确定补充产品，这些产品与该产品捆绑在一起可能有助于提高收入。寻找当前在社交网络(例如，Twitter)中传播的哪些帖子很可能变得非常类似于，比方说#Pizzagate的传播痕迹，这是预测查询。这类帖子是假新闻的最佳候选对象。如果进一步分析发现它们是假的，我们可以采取干预行动来遏制它们。考虑医疗轨迹数据库(DB)，其中轨迹是对患者的带有时间戳的观察/测量的序列。这项任务是找出数据库中哪些患者在入院后6个月内最有可能需要氧气治疗，这与Top-k对预测的询问密切相关。专家希望找到可以将这种可能性降至最低的干预措施。在这个项目中，我们的目标是开发一个通用的数据框架，用于查询模型中的预测并推荐干预措施。这两个都是新的研究方向，以前没有解决过，肯定会在数据科学和决策方面开辟新的天地。以下是我们的技术的一些可能的实例示例。首先，与现有的工作不同，我们将使用一种新的效用驱动模型来捕捉竞争和互补项目之间的复杂相互作用。对于B)，我们将开发虚假内容的ML模型，通过查询知识图来检查事实，并通过干预和缓解活动来对抗错误信息。对于C)，我们将开发医疗轨迹的预测模型，并通过将假设推理与预测分析相结合来制定建议干预的策略。这些应用程序将影响社会的不同方面：营销、打击假新闻，以及对医疗轨迹的干预。我们的技术将把科学放在应用的前沿和中心。

Advances in machine learning (ML) have fueled many successful applications of predictive analytics and recommender systems. Propelling these successes to the next level calls for prescriptive analytics, with 2 key functionalities: i) query over predictions made by models and ii) recommend intervention actions, which if taken, may lead to desired outcomes as predicted by the models. A major long-term vision of this program is to develop a framework, theory, models, and algorithms to realize these functionalities, instantiated on 3 key applications: A) viral marketing (VM); B) misinformation containment; and C) intervention recommendations over medical trajectory databases. The applications have been chosen carefully so that they are related but diverse, they feature heterogeneous data such as graphs, streams, and sequences, and will serve to illustrate the generality of the models and techniques that we will develop. Given snapshots of an ongoing VM campaign, a marketer would want to know if the outcome is likely to meet the sales target. Further, if the product encounters new competition from rival companies, she might want to identify complementary products, which when bundled with that product could help boost the revenue. Finding which posts currently propagating in a social network (e.g., Twitter) are likely to become very similar to, say the propagation traces of "#Pizzagate", is a prediction query. Such posts are good candidates for fake news. If further analysis reveals them as fake, we can take intervention actions to contain them. Consider a medical trajectory database (DB), where a trajectory is a sequence of timestamped observations/measurements on patients. The task, of finding which patients in the DB are most likely to need Oxygen treatment within 6 months of their admission, is strongly related to a top-k query over predictions. An expert would like to find interventions that can minimize this likelihood. In this program, we aim to develop a generic framework for querying over predictions from models and for recommending interventions. Both are novel directions of inquiry, not addressed before, and are certain to break new ground in data science and decision making. Here are some possible example instantiations of our techniques. For A), unlike existing works, we will capture complex interactions between competing and complementary items, using a novel utility-driven model. For B), we will develop ML models of fake content, fact check claims by querying knowledge graphs, and counter misinformation via interventions and mitigation campaigns. For C), we will develop predictive models for medical trajectories and develop strategies for recommending interventions, by combining hypothetical reasoning with predictive analytics. The applications will impact different aspects of society: marketing, fighting fake news, and interventions over medical trajectories. Our techniques will put science front and center in the applications.