This track is intended to bring together practitioners and researchers who are working on the cutting edge of Operations Research and Advanced Analytics to share new areas, explain open problems, formalize new problem areas that are just coming to the forefront, and define challenging questions for further development. Learn more about confirmed presentations in this track below.
Ten Things You Need for AI-Enabled Knowledge Discovery
Uncovering insights and deep connections across your unstructured data using AI is challenging. You need to design for scalability and an appropriate level of sophistication at various stages in the data ingestion pipeline, as well as post ingestion interactions with the corpora. This session will discuss the top ten things, including the techniques that you would need to account for when designing AI-enabled discovery and exploration systems to augment & assist knowledge workers make good decisions. These include document cleansing/conversion, pre-processing, machine-learned entity extraction and resolution, efficient methods for indexing, knowledge graph construction, natural language queries, passage retrieval, relevancy training, relationship graphs and anomaly detection.
Approaches and Open Questions in Fraud Detection: A Call to Action for the Analytics Community
Fraud has been identified by the Global Economic Forum and other international organizations as one of the greatest risks to the economic stability and continued growth of the world economy, with various estimates of the annual global cost exceeding $4 trillion. The core ideas of this talk are focused on how analytics can have an impact on this problem, the methods that work (and how well), and open and pervasive problems that are yet unsolved. While the talk will include examples of fraud in many industries, such as banking, insurance, healthcare, online retail, and many others, the author will emphasize fraud in the public sector based on his personal experience. We discuss the characteristics of fraud behavior and how these can be modeled, and we will explain differences in identifying, interdicting, stopping, or recovering money stolen depending on the stage of the activity. We will also provide a description of the levels of organizational maturity (awareness, data availability, knowledge of methods and tools, and processes and procedures in place) that can be used to differentiate the likelihood of an organization’s success in combating fraud. The talk will also contain details of analytical approaches focused on different stages of an organization’s maturity (and data availability) including rule-based systems (which are often the only tools available in early stages of an organization’s anti-fraud maturity) and more advanced analytical tools including machine learning, supervised and unsupervised data mining, natural language processing in social media spaces, and network-based investigative tools. Finally, we will discuss a number of open problems, including the challenge of emerging fraud (the new fraud approaches that are constantly being developed), including strategies that have been used and research done on this very challenging problem.
Minimizing Risks through Data Driven Decisions
Data is at the core for making decisions. With digitization completely transforming the landscape, businesses are collecting more data than ever before and use it to constantly make better informed decisions to assess risks. In turn, this is disrupting industries globally. Join this session, where we will share experience gleaned from our customers, and extrapolate to bring key best practice learnings: how to minimize risks through data driven decisions; where technology and products like Azure Synapse, Azure Arc, Azure SQL Database Edge and SQL Server 2019 Big Data Clusters can play a role, and develop a vision to build a continuous intelligent world, reducing business risks all the way from the ground to the cloud.
Adversarial Machine Learning
Modern machine learning systems are susceptible to adversarial examples; inputs that preserve the characteristic semantics of a given class, but whose classification is incorrect. Current approaches to defense against adversarial attacks rely on modifications to the input (e.g. quantization) or to the learned model parameters (e.g. via adversarial training), but are not always successful. This talk will include: 1) An overview of attacks on machine learning (data poisoning, model poisoning, adversarial examples) and defenses. 2) Discussion of the enablers of successful adversarial attacks via theory, and empirical analysis of commonly used datasets. 3) Discussion of recently proposed defenses that change the representation of the model outputs, drawing upon insights from coding theory. 4) Novel approaches to detection of adversarial examples using confidence metrics
Explainable Artificial Intelligence as a Tool to Understand What AI Algorithms Actually Do
There have been some amazing improvements to Artificial Intelligence (AI) Agents, with companies like DeepMind showing AIs outperforming Go masters, playing video games at an expert level, and even IBM’s Watson being deployed in medical decision-making. Despite these advances, there are some core features of many of the most popular AI algorithms which make their decisions effectively black box. In the context of Go and video games it may be appropriate to anthropomorphize these AI agents’ goals, however, when these algorithms are being used to determine whether an inmate is paroled or whether an automated weapon’s platform decides to take a kill-shot, then we need to understand how the system makes the decisions that it makes. In this talk, I will review what some popular algorithms are actually learning, how the nature of the data you input influences the output of the system, and present some recent advances in introspection algorithms which work synergistically with AI agents to support and interpret their decision-making capabilities.