A Journey Through The Far Side Of Data Science
- 1. A Journey Through The Far Side of Data Science Ted Washburne @ Stevens Institute of Technology / Sept. 27, 2018
- 2. Your competition is pursuing opportunities for AI & smart automation offers in every business unit and going beyond cost savings or productivity improvements The MBA is no longer more valuable that a BA degree in computer science with machine learning expertise SalesForce asserts that “66% of a sales rep’s time is spent not selling” and AI can automate the drudgery work like onboarding The younger wealth now have 24/7 service expectations that can only be delivered by intelligent automation Driving revenue growth and profitability Lowering the cost of many financial activities to near-zero by understanding the critical building blocks in designing an AI and smart automation strategy Data alone doesn’t help business leaders transform their organizations. AI enables faster and larger-scale intelligent process optimization, intelligent agents, and innovation 1 Data is the new foundation 2 AI becomes the new norm 3 Innovation is intensifying 4 AI for the Front-Office 5 Technology is the Business 6 AI enabled Customer Experience 7 Everything happens in the Platform Reach and depth of AI technology is transforming Business
- 3. … Intelligent Automation (Aitomation) is the key … Complete a thorough benefit analysis before committing to investments Make aitomation a strategic imperative and get senior leadership backing Focus on AInnovation with a central team providing Governance Identify and rapidly scale high- impact aitomation use cases Artificial General Intelligence is making significant inroads in Legal and Accounting Develop the people capabilities needed for maximum value (Python, Machine Learning) Honestly asses the Competition and establish ecosystem partners to challenge them
- 4. Who Made this all Possible? Paul Werbos & The Chain Rule • https://en.wikipedia.org/wiki/Ba ckpropagation • https://en.wikipedia.org/wiki/Pa ul_Werbos • http://explained.ai/matrix- calculus/index.html • differential matrix calculus, the shotgun wedding of linear algebra and multivariate calculus. Silicon Graphics -> Nvidia + Ian Buck • Inventor of CUDA, the established standard for GPU Computing worldwide. Built engineering team from two people into international and matrixed organization in numerical libraries, compilers, system software, IDEs, profilers, debuggers, APIs, QA, build and release, and AI frameworks. https://www.nvidia.com/en-us/deep-learning-ai/education/
- 5. Data Science Themes at SIT Artificial Intelligence, Machine Learning & Cybersecurity "We knew it was the Russians, and they knew we knew," Johnston told the NYtimes of the cyberwarfare. "I would say it was the cyber equivalent of hand-to-hand combat." Biomedical Engineering, Healthcare & Life Sciences Deep Learning’s Deepest Impact: AI Storming Through $6.5 Trillion Healthcare Industry Complex Systems & Networks USAA developed a comprehensive analytical simulation model of their complex operations and the market that is allowing senior executives to explore a wide range of scenarios and strategic options and understand long term implications of these decisions. Data Science and Information Systems Cognitive knowledge graphs encode a model of expert knowledge of every domain within a context. This gives bots a semantic understanding of the context and helps them to respond to complex queries. Financial Systems & Technologies Providing an extension of the bank's main quant team covering the whole range of quant tasks from numerical algorithms to multi-curve building and financial modelling. Resilience & Sustainability Utilities now use machine learning to classify network assets at high risk of failure and manage the complexities of distributed energy resource management. Preparation for cyberattacks include Transformers, circuit breakers, as well as secure warehouses to store them in select locations and the preplanned transportation and logistics to get them where they need to go as the situation requires. Maritime Security HawkEye 360, a developer of space-based radio frequency (RF) mapping and analytics systems develops deep convolutional neural networks, Bayesian propagation networks and statistical anomaly detection for maritime domain awareness Systems Engineering Research Center *Constraint Programming to Incorporate Engineering Methodologies into the Design Process of Complex Systems. *Recurrent Nets to solve Knapsack problems in Mission Management. *Petri Nets for Process Discovery https://www.usnews.com/news/articles/ 2016-09-23/is-the-energy-grid-in-danger Chua, L.O., Lin, G.-N “Nonlinear programming without computation”
- 6. Early Years
- 7. Career
- 8. SIT Courses & Case Studies •Citi Cards •Capital One FIN 615 Financial Decision Making •Ceph •Schneider Electric •Spark MIS 630 Database Systems and Decision Support •Snowflake •Merging Accounts •Bad Days MIS 636 Data Warehousing and Business Intelligence •Knapsack •HSBC RPA BIA 650 Process Optimization and Analytics •BMW China •Coin-OR •Visa Fraud •Morgan Stanley High Frequency BIA 670 Risk Management: Methods and Applications •Kaggle •Numpy •Pandas •Seaborn, Keras BIA 652 Multivariate Data Analytics •Schneider Electric •Unmanned ground vehicle • Capital One BIA 664 Data and Information Quality •TV Advertising •Direct Mail BIA 654 Experimental Design •MBA Forecasts •Asurion Trouble Prediction MIS 637 Knowledge Discovery in Databases •DBS Treasury •Hastie, Stork BIA 656 Statistical Learning and Analytics •Flu prediction •Amobee link prediction •Alternative Influence Network BIA 658 Social Network Analytics •DoubleClick Attribution •Yelp reviews BIA 660 Web Mining •PwC Audit.ai •Automated Feature Engineering BIA 662 Cognitive Computing •A template for understanding Big Debt Crises •by Ray Dalio BIA 670 Risk Management: Methods and Applications •Omnicom •HP Forecast •Lowes Hardware •Cambridge Analytica BIA 672 Marketing Analytics •Walmart Labs •Nike •NV Energy •Hawkeye360 BIA 674 Supply Chain Analytics •DBS ATM Maint. •Insurance Telematics •ASW BIA 676 Data Streams Analytics: Internet of Things •Spark killed Hadoop •MapReduce is dead •DataBricks •SparkFlows.io BIA 678 Big Data Technologies •Next Best Action Models for Wealth Management •HR systems BIA 686 Applied Analytics in a World of Big Data https://www.youtube.co m/watch?v=g6oIQ5MXBE4 https://aws.amazon.com/rekognition/
- 9. Useful Algorithms For Your BI Career Reinforcement Learning Dueling Deep Q Network Robots for unloading ships, warehouse forklifts, harvesting crops Recommender Systems Netflix Prize A quantum- inspired classical algorithm for recommendation systems. Ewin Tang, July 10, 2018 Segmentation Decision Tree Clustering (k- means or EM) & factoextra •Avoid “lazy” dimensionality reduction, like principal components Forecasting Tsintermittent Recurrent NN XGBoost •http://docs.h2o.ai/ driverless- ai/latest- stable/docs/usergu ide/time- series.html?highlig ht=forecast Regression GLM/GBM Classifiers Trees or XGBoost Logistic Regression Convolutional Neural Nets NLP Word2Vec Truncated SVD OR Coin-OR branch and cut Knapsack DS is IT + Biz + software engineering + design Get to know Docker, Kuberflow, Kubernetes, and SeldonA 2018 paper in Nature cited AlphaGo's approach as the basis for a new means of computing potential pharmaceutical drug molecules
- 10. Systems Engineering Research • I authored a paper on a Hopfield (recurrent) system that we had built at Lockheed Research Labs, along with a few other contributors and got it accepted to a Neural Networks conference in Maryland – as a poster paper • Some weeks later, I am getting a new clearance and told to report to a building with no windows. • The project has something to do with Mission Management and it has a challenge in the area of solving a really big knapsack optimization problem over multiple time windows • The problem often arises in resource allocation where there are financial constraints and is studied in fields such as combinatorics, computer science, complexity theory, cryptography, applied mathematics, and daily fantasy sports. • While researching the problem, I was concerned about a limitation of Hopfield networks getting stuck in local minima and not finding the globally optimal solution • Researching all the way back to the 1970’s, I found a paper by a Berkeley professor named Chua that had a better behaved network for this problem • The network could be modeled in the popular circuit modeling software “SPICE”, using a Sun SPARC workstation, like was done at Analog Devices and Linear Technology (merger last year) • SPICE (Simulation Program with Integrated Circuit Emphasis) is a general-purpose, open source analog electronic circuit simulator. MatLab can do this now • We found that we could set up very large constraint matrices and completely random starting points and the system would find an almost perfect solution in under a second and then spend the next 10 seconds converging to the optimal solution • Far faster than a Cray using the traditional linear programming ‘Greedy Algorithm’
- 11. Maritime Security • Neural Nets for Sonar Signal Processing
- 12. Financial Systems & Technologies NBA: Recommended Solution for Client Investment & Trading Tech / Treasury and Markets Sales Person Counterparty Client Product Counterparty Market Data Quantitative Product Info Market Data Revenue P&L Records Unstructured Data (WWW) Counterparties clustered by similarity (clients and non-clients) Graph or relations • Salesperson x CPTY as nodes • Products related to each graph edge Temporal structure of relations: Mutations of the graph in time Market Regime Switches • Regime-driven behaviour (SP, CPTY) • Regime-driven revenue and risks Contracts/products parameters templates for each SP/CPTY in time Client Request Historical Data Knowledge Discovery OUTCOMES •Stable transitions? •Periodic events? •Hidden connections? •Recognized states? •Unstructured outlier? •Atypical states series? •Unrecovered gaps? •Unstable behaviour? Patterns Anomalies Client Existing Data Required Extra Data for Future Phases Discovered KnowledgeKey: NBA Data NBA Model MODELS
- 13. Data Science and Information Systems • https://databricks.com/session/moving-ebays- data-warehouse-over-to-apache-spark-spark-as- core-etl-platform-at-ebay • Snowflake - Beyond Hadoop: Modern Cloud Data Warehousing • Captial One / NY / Sr. Software Engineer • All of our infrastructure runs on AWS, and we are eyeing other cloud providers too. We use Elasticsearch and ELK stack, Redis, PostgreSQL, Redshift, and Snowflake. We do analytics using H2O, Spark and MLlib, Databricks/EMR, TensorFlow and Keras. We build awesome products for our users that use this data. We write microservices in Go and Node.js, orchestrated by Kubernetes, with user experiences written in React and TypeScript. We embrace serverless. Your hands-on expertise in at least some of these tools will be valuable, as well as your track record of providing effective technical guidance.
- 14. CPA + AI Accounting jobs are not going away – the skill set is changing • Business Setting • PwC partnered with H2O.ai to build a revolutionary bot that uses AI and machine learning to ‘x-ray’ a business, analyzing billions of data points in milliseconds, seeing what humans can’t, and applying judgement to detect anomalies in the general ledger. Called GL.ai, it is the first module of PwC’s Audit.ai. • Approach • GL.ai harnesses PwC’s global knowledge and experience, embedding it in algorithms trained to replicate the thinking and decision-making of expert auditors. • It examines every uploaded transaction, every user, every amount and every account to find unusual transactions (indicating potential error or fraud) in the general ledger, without bias or variability. • Impact • Experience confirms that GL.ai speeds up the audit process, generates insights that boost efficiency, and provides comfort that attention is being focused on areas of true risk. These benefits are a direct result of GL.ai’s ability to analyze huge amounts of data, not limited by sampling. • The next Audit.ai modules are in development. They are set to revolutionize the audit, enhancing client service, quality and efficiency, and giving our people more time to do what machines can’t: thinking strategically and engaging, communicating and building the relationships needed to turn data insights into business action.
- 15. Complex Systems & Networks Prescriptive Models combine: •Known system facts, structure, and process •Facts derived from data and statistical analysis •Causal hypotheses (business judgment and assumptions) •Dynamics of the system •To not only predict the behavior, but to also to tell you why the system behaves that way •Point to actions you can take today •Prepare for responses if certain events materialize (real options) •Help to point out wrong assumptions •Explore a wide range of outcomes (scenarios) Hybrid Modeling •Discrete Event Simulation – Models how entities flow through a process and consume resources. Good for finding bottlenecks and throughput/volume issues for well defined processes and non-adaptive entities (PROCESS CENTRIC) •System Dynamics (Causal Loop) – Models aggregate causal relationships between system components to study system level behavior. Incorporates non-linear causation, feedback loops, delays, system interdependencies, and soft variables. Good for broad, system level understanding of dynamic behavior (SYSTEM CENTRIC) •Agent Based Modeling – Models individual agents and how they react to external stimuli and their relationships with other agents. The complex dynamic system level behavior emerges from the interactions of simple agents following simple rules. Agents have biases and bounded rationality. They adapt and learn, but operate in a noisy uncertain environment (INDIVIDUAL CENTRIC) Superior Models •Far Higher Granularity, Time •Skills/Groups, Missing Effects •Abandonment behavior, Multi-skilled Sales Reps, Routing logic •Also Incorporate Call and Sales Rep Attributes, Attribute based routing, Individual Agent Behaviors Business Results •Benefits •Increased Revenue due to reduced abandons •Improved customer satisfaction due to reduced wait times •Reduction in hiring and training costs •Risk Reduction •Better understanding of operational risks and where they might surface •This allows USAA to design mitigation strategies that are more proactive than reactive •Investment Prioritization •Reduced/Avoided rework due to better sequencing of work •Better allocation of resources to create most value How this Relates to Strategy and Value 8 EXTERNAL FACTORS LEADERSHIP DECISIONS PROCESS OUTCOMES/ MEASURES (KEIs)Competition GDP Capability Investments Hire Staff Customer SAT Quality Profit contacts apps products rework VALUEOperational ProcessesStrategic Choices Risks Causal Hypotheses -Analytics -Business Judgment -Assumptions Example II – Portfolio Roadmap and Dependencies 15 * - Notional Data Incorporate data analytics as well as business judgment in your models Using Data Science and Simulation to Create Business Value Dr. Bipin Chadha - Data Scientist USAA Enterprise Data Analytics Office Nov. 2015
- 16. Biomedical Engineering, Healthcare & Life Sciences Best Opportunities to Healthcare’s Needs Clinical • Value Stream Mapping/Design, Leading Kaizen and Relentless Root Cause Analysis • Personalization of care using claims and biometric data (Apple watch?) models • Predict outcomes and adverse events • Accountable care – Medical Economics Enterprise • Cybersecurity • Payment Integrity • Pricing Optimization • Fraud, AML, KYC • Intelligent RPA with Petri Nets for Process Discovery Marketing & Sales •Rep-Broker-Sponsor Attribution Modeling •Real-time targeting in the call center •CLV, health behavior/activation models, experiment design •Attribution models •Develop audience segmentations, core value propositions, messaging strategy for the different segments, measure impact and efficacy of marketing investments •End to end delivery of behavior change campaigns •Chatbots & knowledge graphs •Customer Journey Analytics •Direct-to-Consumer marketing & sales Member Experience • Advocacy Analytics • Flu region prediction (with active listening of Social Media) and Next Best Action messaging • Churn prediction • Personalized Robo-advisors provide greater convenience and insightful, real-time recommendations
- 17. Artificial Intelligence, Machine Learning & Cybersecurity Detecting Fraud or Cybersecurity Transactions Involves Monitoring Multiple Views Simultaneously Endpoint authentication Is the session being compromised? Is this accounts' behavior normal for this channel? Is this accounts’ behavior normal for all channels? Are multiple accounts behavior showing correlated behaviors across multiple channels? Endpoints Merchant POS ATM Online Purchase Acquirer third-party online payment platform Cloud Communication Network Production Fraud Scoring •Endpoint authentication •Is the session being compromised? •Is this accounts' behavior normal for this channel? •Is this accounts’ behavior normal for all channels? •Are multiple accounts behavior showing correlated behaviors across multiple channels? Transaction histories with fraud categorization •Machine learning fraud models •Cross account and cross channel graph analytics Issuers Banks Alternative Payments Enablers Blockchain & Bitcoin
- 18. A Unified Approach to Interpreting Model Predictions Scott Lundberg, Su-In Lee
- 19. What Capital One is Looking For FINANCE ASSOCIATE Finance • Analyze financial metrics and performance • Develop, improve and / or automate reporting and analysis to provide insight into business trends • Play a key role in evolving product and strategy decisions by providing finance analysis and forecasts • Prepare for the future by learning new Tech skills Accounting • Participate in the external financial reporting process, including the quarterly earnings release and securitization trust reporting • Perform financial and operational audits, testing controls and identifying efficiencies • Monitor/Enhance business specific analytics in support of external and internal financial reporting • Evaluate and engage in Robotics Process Automation (RPA) projects across the Controller’s Organization. Preferred Qualifications: • Bachelor’s degree in Finance, or Economics, or Business, or Accounting • A demonstrated interest in financial management and technology aptitude • At least 6 months of experience or course work in Financial Planning & Analysis (FP&A) • Aptitude with technologies such as Python, SQL and R is strongly preferred DATA SCIENTIST INTERN On any given day, you might: • Evaluate open source and internally-developed modeling and analytics tools using real business data • Integrate internal data with external data sources and APIs to discover and implement actionable insights • Design and craft rich data visualizations to communicate stories to customers and company leadership We'd love to find someone who is… • 1. Intellectually curious. You ask why, you explore, and you are excited to imagine and create new ideas by inventing self-adaptive models or by tapping into unstructured data sources. You love mining data for insights into behaviors, intent and sentiment. • 2. A builder. You are passionate about delivering better experiences and better products to our customers and have a deep sense of ownership for your craft. • 3. An experimental scientist. You love putting on your lab coat and trying new things, new combinations of tools, techniques and feature engineering approaches even if you sometimes fail. Basic Qualifications: • - At least 6 months of experience or course work in open source programming languages for data analysis • - At least 6 months of experience or course work in inferential statistics or machine learning Preferred Qualifications: • - Direct experience with either Python or R, plus one other general purpose programming language such as Java or C/C++ • - Experience or course work with large scale data analysis
- 20. Further Reading Worthy Of Becoming The Basis Of Your Thesis • https://amp-businessinsider- com.cdn.ampproject.org/c/s/amp.businessinsider.com/why-attitude-is- more-important-than-iq-2017-2 • https://www-theverge- com.cdn.ampproject.org/c/s/www.theverge.com/platform/amp/2018/9/5/ 17822562/google-dataset-search-service-scholar-scientific-journal-open- data-access • https://www.liebertpub.com/doi/full/10.1089/big.2018.0083 • https://blogs.microsoft.com/blog/2018/07/19/powering-our-customers- the-innovation-story-behind-microsofts-earnings/ • https://www.fastcompany.com/40590772/my-three-decades-at-disney- taught-me-not-to-fear-automation
- 22. Appendix H2O Materials Cheat Sheets for AI, Neural Networks, Machine Learning, Deep Learning & Big Data
- 23. H2O Materials Publically Available Materials http://docs.h2o.ai/ https://github.com/h2oai/h2o-meetups
- 32. Cheat Sheets for AI, Neural Networks, Machine Learning, Deep Learning & Big Data https://becominghuman.ai/cheat-sheets-for-ai-neural-networks-machine-learning- deep-learning-big-data-678c51b4b463