Full Mega-PAW 7-Track Agenda – Detailed Session Descriptions

Understanding the intraday microstructure dynamics across different universes of stocks and markets is fundamental in designing any optimal  trading strategy, especially for trading diverse portfolios of stocks. In this talk we discuss how modern machine learning techniques can be used in conjunction with dynamical modeling of the intraday phenomena to identify those trading strategies that move away from the general country/sector classifications but rather respect the stock's particular microstructure characteristics.

Bringing the benefits from AI efforts to the frontline workers continues to be a struggle across major healthcare organizations.  We worked on a novel, practical approach to directly take on the workflows of healthcare workers.  This session shares the successes and failures in our attempts, and the AI's introduced via this approach to achieve efficiency and/or outcome goals. This practical workflow approach uses AI as tools, hence can deploy various AI’s for a variety of problems including patient status tracking and task automation. AI's being directly in the workflow also enables continuous learning, process improvement, and optimization toward specific goals.

This project sets forth the work to be performed for the Customer Acquisition Model, aiming to score the entire through-the-door (TTD) populations and eventually make optimal lending decisions. The work will be divided into two phases: Phase 1 will concentrate on building a Minimum Viable Product (MVP). This phase will leverage existing techniques used currently, where applicable, including the target, data sources and transformations. Phase 2 will conduct another round of data exploration with the purpose of identifying additional transformations to increase model lift over what was already achieved in Phase 1.

Until recently, healthcare has not understood root causes of diseases well enough for prevention; the main approach has historically been to treat patients after onset. While primary prediction scoring systems are routine for CVD patients, the goal is to reach patients before primary events occur. Amgen and a startup partner are co-developing a machine learning solution that uses existing EMR data to develop statistical and machine learning models predicting secondary CVD events. Having more accurate risk prediction models could significantly impact approaches to disease prevention. Session will also cover role of partnership in sourcing, prototyping, piloting, and scaling novel technologies.

Many organizations are faced with the challenge of how to analyze their sensitive data without hosting it on any public cloud. This talk will focus on companies who collect data from their factory operations and are interested in predicting mechanical failures. The audience will get an overview of how to formulate their business problem, perform feature engineering and build a predictive maintenance model using R/Python.

Field issue (malfunction) incidents are costly for the manufacturer’s service department. A normal telematics system has difficulty in capturing useful information even with pre-set triggers. In this session, Yong Sun will discuss how a machine learning, deep learning based predictive software/hardware system has been implemented to solve these challenges by 1) identifying when a fault will happen 2) diagnosing the root cause on the spot based on time series data analysis. Yong Sun will cover a novel technique for addressing a lack of training data for the neural network based root cause analysis.

Many companies regardless of their size and years in business may not actually have an analytics team or may have a team of one. During my last speaking engagement at PAW I spoke with a great deal of folks who were interested in creating analytics units but didn't really know how to go about it or were under assumption that it would be very cost prohibitive. This is a case study of a team that consists of former spreadsheet guy, grad with a fresh masters in engineering, a former rocket scientist and two former workflow coordinators.

A frequent criticism of the use of machine learning models as compared to human analysis is that ML models are "black boxes" and uninterpretable. Recent advancements in the field of explainable AI allow us to understand what factors influenced both individual predictions and aggregate model behaviors. We will revisit a case study from another PAW conference on predicting hospital readmissions, except this time we will use open-source software and dive into the 'why' with various visualizations that explain the model's behavior.

We will discuss the approach used to learn topics and their place in a multi level hierarchy on hundreds of millions of text records. These methods are generalizable beyond the domain in which they were applied. We used a combination of supervised and unsupervised machine learning methods, which we will discuss at more length including the technologies, algorithms, and results.

In this session, I will provide an overview of logistic regression, GLM logistic regression, decision tree, random forest, gradient boosting, neural networks, etc. Then, a comparison will be made through a case study about building a full life cycle of predictive model based on insurance datasets. Since the business goal through the feature engineering stages are similar given the same case study, the comparison of each method, its  advantages and disadvantages, will include the feature selection, model building, model validation, and model testing stages. Also, model implementation and interpretability will be discussed and compared. Finally, we will discuss implementation of these methods in Python and R.

Emergency departments have seen a dramatic increase in the number of visits from elderly patients. Many elderly use a personal emergency response system (PERS) to signal for help in case of an incident such as a fall or breathing problems. At Partners Healthcare, we are testing a predictive model that uses PERS data to predict elderly at high risk of emergency department visits. Clinical staff from our homecare program perform interventions with high-risk patients. This presentation will cover the development of the predictive model and its deployment in a randomized controlled trial.

Industry 4.0 can suffer from a real-world application problem when industrial and manufacturing companies only view IoT as a solution during large-scale plant upgrades or new construction. This case study presents how a manufacturing company has been able to generate energy cost savings in the scale of $MM through targeted deployment of sensors and IoT connected equipment into existing, large, (and sometimes very old and dirty) machinery and factories. By lowering the threshold of what projects are considered worthy of an IoT investment, what was previously considered run-of-the-mill operations can suddenly provide insightful information and an exciting ROI.

Wavelets have long been known for their strong ability in denoising and transforming signals, for example in speech recognition and image processing. More recently, Long Short Term Memory Networks (LSTMs) and Autoencoders started showing promising results in this space as well. Application is possible also to data sets generated by complex nonlinear processes with a low signal to noise ratio. Both Wavelets and LSTMs offer value through generating cleaner training data and ultimately driving deeper insights. In this session, we will show three concrete use cases: Analysis of Financial Time Series, Sales Forecasting and Credit Risk Analysis.

An emerging biotech company launched the first treatment option in an unestablished rare disease market. Extremely low prevalence, lack of physician awareness, no codified ICD-10 diagnosis code, and the lack of approved treatments resulted in significant mis-diagnosis, making the application of AI challenging. Addressing the challenge required combining first and third party de-identified data in a HIPAA-compliant workflow based on Swoop's prIvacy platform. Of the 84 start forms in the past 6 months, 24 (29%) were due to Swoop's AI model. Further validation is underway in a university hospital system, embedding predictions into clinical workflows to improve patient outcomes.

Optoro’s three core data culture problems were the following:

• Fear of Data
• Inconsistent Use of Vocabulary and Metrics
• Data Mistrust

This presentation will outline the strategies that we used to combat these issues. This on-going endeavor is yielding benefits to Optoro, including
• Increased alignment on company goals
• Improved ease of communication between teams and with Senior Management
• Consistency across external messaging

The use of AI in decision making processes brings efficiency and data-driven results, but also risks.  Machine learning creates models which make predictions based upon patterns learned from past data. The reasoning behind these decisions is not available to the users of the models, or recipients dealing with the consequences of the decisions. E.g., a sales person doesn't know why a business is a good lead, and credit doesn't know why credit is denied.  This is a case study on adding explanations to machine learning algorithms, so that users will have greater confidence and insight into machine-driven decisions.

In today’s digital age, users expect a fast, reliable mobile experience. Degradations (also referred to as regressions) in mobile app performance affect not only user experience, but even hurt business metrics. However, existing mobile app release pipelines lack the necessary infrastructure to detect regressions in a mobile app's performance before it is rolled out to the world. At Uber, we are building a state-of-the-art mobile regression detection pipeline, with the goal to detect regressions as small as 1%. Our approach includes both technological innovation as well as employing machine learning along with statistical testing techniques to improve the sensitivity of the regression experiments.

Predictive models are increasingly used for important decisions such as which customers may open a financial account. These decisions affect the opportunities available to customers and drive business results. To maintain the customers' trust, it is important to be able to explain individual predictions. Likewise, it is important to explain the model logic for business managers, compliance professionals and regulators who expect fair decisions.  This is not easy when using advanced techniques such as ensemble models.  Mr. Duke will share Experian's recent advances in explainable AI technologies, with results in credit risk modeling, synthetic identity detection and fraud prevention.

Healthcare has always used statistical analysis and analytic capabilities for accounting, reimbursement, actuarial and fiscal projection purposes. New developments in advanced statistical and predictive analytics techniques promise to revolutionize health and medical outcomes, and care delivery. These new techniques utilize modern machine learning and Artificial Intelligence methods to predict and prescribe at the individual level, instead of using traditional statistics. Learn how new machine learning techniques are being used for value-based purchasing, population health, healthcare consumerism and precision medicine. Peer into the future of Healthcare Data Science with predictions from industry leaders.

"Build a better mousetrap and the world will beat a path to your door." Build a dozen mousetraps, with different triggers, bait, and alarms all meant for different species of mice, and the world will beat your door down. This presentation addresses the challenges of alert fatigue generated by successful predictive models. When your target audience is presented with multiple models recommending different and occasionally overlapping but important actions, there needs to be harmony in the message sent. This presentation examines the interactions of prioritization, frequency, severity development, consistency, messaging, and user socialization across multiple models for effective action.

How much data is enough to build an accurate model?  This is often one of the first and most difficult questions to answer early in any machine learning project.  However, the quality and applicability of your data are more important considerations than quantity alone.  This talk presents some insights and lessons learned for gauging the suitability of electronic health record (EHR) training data for a desired project.  You will see how to determine if more data might increase accuracy and how to identify any weaknesses a model might have as a result of your current training data.

While many businesses understand the value of advanced analytics in decision making, operationalizing data science can be challenging. This session reveals how Enova International has been successful at integrating traditional operations with advanced analytics to turn fraud defense into a collaborative analytics function. Over time, through combining the latest technologies in data, machine learning and decision automation with manual investigations, Enova International has been able to attract and retain top analytics talent, mitigate fraud risk, improve profitability and deliver a better customer experience.

With the advent of big data and machine learning, there is an opportunity to combat rising healthcare costs by leveraging data in an ethical and privacy compliant way to establish more consistency and implantation of preventative care. We need to ensure there is a fundamental set of rules and responsibilities in place among healthcare organizations to protect their patient's privacy. In this presentation we will address this challenge and speak to the importance of creating an ethical and privacy compliant approach to aggregating multiple data sources which then can be used to improve patient outcomes.

Vistra Energy is one of the largest energy companies in the United States, owning both power generation and retail operations in extremely competitive markets throughout the country. We combine the big data opportunities we have available as a utility with advanced analytics to provide premium customer services that differentiate us from other utility providers. We will present three case studies demonstrating how we are able to leverage our firehose of 15-minute interval IOT device energy usage data from over 1.5 million customers with advanced modeling techniques to provide added value to our customers and increase brand loyalty.

Predictive analytics has been a buzzword for a few years now. It has seen success for a wide range of applications. Within the Insurance industry, several applications have emerged, across underwriting, claims, marketing, and beyond. In this session, we will highlight examples of success factors that help sustain predictive analytics in a workers compensation insurance environment.

Independent Pediatricians typically maintain daily patient volumes of 20-30 patients to keep their practices viable. Pediatricians also schedule appointments up to a year in advance, leading to as many as 15% of patients not showing up for appointments each day. The financial and clinical impact of these gaps in pediatric appointment books is substantial.

PCC and Rexer Analytics analyzed pediatric no-show patterns to identify the variables that truly affect appointment truancy. These insights were translated into interventions to reduce patient truancy. We present pediatric no-show patterns, key predictors, and the results several Pediatric practices are seeing with targeted interventions.