Microsoft Data Platform Airlift 2017 Rui Quintino Machine Learning with SQL Server 2016 and R Services

Editor's Notes

• #5: Basic definition: Machine learning develops algorithms for making predictions (statistical sense) from data * Learning models from available training data, to make good predictions on unseen test data
• #6: Basic definition: Machine learning develops algorithms for making predictions (statistical sense) from data * Learning models from available training data, to make good predictions on unseen test data
• #7: Confidential… + notes From the book: AzureMachineLearning – AzureFundamentals Many examples of predictive analytics can be found literally everywhere today in our society: Spam/junk email filters These are based on the content, headers, origins, and even user behaviors (for example, always delete emails from this sender). Mortgage applications Typically, your mortgage loan and credit worthiness is determined by advanced predictive analytic algorithm engines. Various forms of pattern recognition These include optical character recognition (OCR) for routing your daily postal mail, speech recognition on your smart phone, and even facial recognition for advanced security systems. Life insurance Examples include calculating mortality rates, life expectancy, premiums, and payouts. Medical insurance Insurers attempt to determine future medical expenses based on historical medical claims and similar patient backgrounds. Liability/property insurance Companies can analyze coverage risks for automobile and home owners based on demographics. Credit card fraud detection This process is based on usage and activity patterns. In the past year, the number of credit card transactions has topped 1 billion. The popularity of contactless payments via near-field communications (NFC) has also increased dramatically over the past year due to smart phone integration. Airline flights Airlines calculate fees, schedules, and revenues based on prior air travel patterns and flight data. Web search page results Predictive analytics help determine which ads, recommendations, and display sequences to render on the page. Predictive maintenance This is used with almost everything we can monitor: planes, trains, elevators, cars, and yes, even data centers. Health care Predictive analytics are in widespread use to help determine patient outcomes and future care based on historical data and pattern matching across similar patient data sets.
• #8: Slide objective Establish that R is a language is as important for the community that uses it an the capabilities written to extend it than the language itself. Talking points Part 1 of the R World is The R language, developed specifically for data analysis – particularly among statisticians and mathematicians. [optional points]: Developed in New Zealand, release in roughly 2000. Maintained by the R Foundation which releases new editions of R every few weeks. Licensed under GPL open source license. R directly supports complex data manipulation operations making them extremely simple for users, particularly those with greater depth in statistics and mathematics than in computer science. Huge community of users across industry, government and academia use R daily. There are R user groups in most major cities. Some of them very active and very large. Suggest that users look at MeetUp for local groups that meet regularly. Most important to the value of R is the huge repository of freely exchanged, algorithms, techniques, scripts, adapters, techniques, training available. Introduce CRAN: “The Comprehensive R Archive Network”. Data access & integration Data transformation Data profiling Data visualization Predictive analytics Machine Learning CRAN contains over 7000 (and growing) contributed packages. Many algorithms, test data, comments on usage, etc. One package contains hundreds of algorithms packaged as a library. All are designed to run with the R language. CRAN is the largest but not the only. Thousands of additional algorithms, visualizations and tools are available from BioConductor, GitHub and other repositories. Notes
• #9: Demo Power BI Desktop. Demos are available at //BI.
• #14: Slide objective Introduce how the use of open source R for machine learning and advanced analytics has been limited to a narrow user base of data scientists. Related to this, also discuss how many challenges and complexities remain for advanced analytics in the marketplace. Talking points Today, advanced analytics using open source R are being performed only by highly trained and specialized data scientists, mathematicians, and analysts who can create and nurture these models. This means that many challenges and complexities remain in the marketplace. First, many companies cannot negotiate the increasing costs of specialized talent, infrastructure, and machine learning tools that make total cost of ownership (TCO) and return on investment (ROI) uncertain. Second, siloed and cumbersome data management restricts access to data and poses limitations on what data can be included in models. Third, trying to collaborate across complex and fragmented technologies tends to limit agility and reduce participation in exploring data and building models. People end up struggling with the technology instead of focusing on the business problem at hand. Finally, many models never achieve business value because it’s so difficult to deploy them to stable production environments. If you can imagine spending hundreds of thousands of dollars on a solution and having it never go into production, you can see why machine learning has been so niche up to this point. Notes
• #18: Key Message: Products from Revolution Analytics are continuing and Microsoft is adapting and expanding platform coverage. Revolution R Enterprise product continue as R Servers New integration of RRE into SQL Server as SQL Server 2016 R Services Revolution R Open continues as Microsoft R Open Additional versions and editions coming for various unique communities – desktop developers, MSDN members, education community RRE / R Server available as the R options for various cloud offerings (CIS, Azure Linux, Azure HDInsights Hadoop, Data Science VM, etc.) Support for rev’s of existing versions of Hadoop and full support for HDInsight in Azure.
• #19: Microsoft R Server is a broadly deployable enterprise-class analytics platform based on R that is supported, scalable and secure. Supporting a variety of big data statistics, predictive modeling and machine learning capabilities, R Server supports the full range of analytics – exploration, analysis, visualization and modeling Slide objective Introduce the high –level value of R Server and R Services over other instantiations of the R language. . Talking points R Server products provide an enhanced experience for the R User without loss of compatibility R Server products are “open core” – the utilize the open source R product entirely and build new capabilities around that core without impacting compatibility. Users of R Server products enjoy full compatibility with open source compatible with the entire (and vast) collection of algorithms, connectors, visualization tools shared openly via CRAN, Bioconductor and other shared resources like GitHub. Key extensions enable R to tackle big data challenges that exceed the capacity of open source R. R Scripts built for one platform using R Server can be easily run on another platform running R Server We call it WODA – write once, deploy anywhere. Two key contributions: Build on any version of the product and deploy using other versions Investment protection as platform choices change Develop on the desktop and immediately deploy to RDBMS – SQL Server, EDW (SQL Server & Teradsata) or Hadoop (Microsoft, Cloudera, Hortonworks and MapR) Notes
• #20: Slide objective Introduce the three value proposition pillars of SQL Server 2016 R Services. Talking points SQL Server 2016 R Services brings the perfect mix of fast querying and In-Memory OLTP optimization from SQL Server 2016, as well as data exploration, predictive modeling, scoring, and visualization from the R Services family of products. It delivers unprecedented enterprise speed and performance for advanced analytics, thanks to near-database analytics and parallel threading and processing. It also delivers scalability and choice not seen before from a stable, commercial platform for advanced analytics. Its on-premises, cloud, and hybrid benefits, as well as its limits with large datasets, are unmatched. Finally, there is no additional cost because the offering is included in SQL Server 2016. In addition, the ability to reuse existing R code and eliminate data movement across machines provides significant value. Notes
• #23: Slide objective Illustrate the potential scale benefits possible with R Server’s ScaleR algorithms. Show a representative example and explain the 3 mechanisms that help achieve the improvements. . Talking points We tested the improved data and computational scale of the R Server’s ScaleR library of enhanced, parallelized algorithms. This is an example. Speed: On a 4 core laptop, with 8GB of RAM, open source R could process about 300,000 events in a particular data set prior to exhausting available memory. The test tool about 77 seconds to run the most commonly used R linear regression algorithm called “lm”. We than ran the same test using our parallelized, rewritten (in C++) linear regression module called rxLinMod. Data Scale Algorithms in the ScaleR library are also rewritten to analyze data in “chunks” to eliminate the memory-limits of typical open source R algorithms. Where the open source lm exhausted memory at about 300,000 events, the improved rxLinMod was working fine at 5 million events where we stopped testing. The result is a 50x performance improvement over open source linear regression, and no memory limits. Parallel Scale This example shows only the effects of running optimized, compoiled code on all cores of a laptop. Greater benefits are available. What is not shown, is that the work done to parallelize across 4 cores can also be utilized to scale across many nodes in systems such as EDWs and Hadoop. While results vary, the system, as you can see, responds linearly with respect to data size. Rehosting using R Server for Hadoop can provide even more dramatic speed and scale results. Notes
• #24: …. In the multi-platform world of on-premises… Slide objective Differentiate R Server from other R offerings such as vendor-specific offers from Oracle, HP, SAP and Teradata Clearly communicate two benefits – develop on multiple machines, and protect investments from platform change disruptions later. . Talking points R Server makes your data scientists’ jobs easier. By running identically on multiple platforms, users can build on one platform, the move the scripts to another. This has several advantages: Run modeling algortihms on systems here larger compute or data storage is available. It also permits modesl to be built on one platform, but model scoring or operationalization to take place elsewhere. Finally, with the availability of very low storage and compute costs in the cloud, users can load, transform, visualize, study and model data in the cloud, but actually run the model computations on on-premises systems. Perhaps more importantly, however, portability across systems protects organizations investments in R-based data science. The best best big data storage and compute platform for today may not be the best choice tomorrow. Compatibility across systems brings the possibility to avoid disruptive recoding efforts when such changes occur. Notes Two ways to underscore this point are to describe the range of compatibility available with other vendor’s R versions. Oracle R, because it works only with algorithms running on Oracle Database, is not portable. The same is true for R implementations from Teradata, HP Vertica, SAP and others. They are in essence platform specific. Another way to describe the problem is with a fictitious story: Imagine a CIO who has to tell his data science team “we’re changing platfomrs, and therefore, you need to change all your programs and scripts to work with the new platform”.
• #26: Demo Power BI Desktop. Demos are available at //BI.
• #27: Basic definition: Machine learning develops algorithms for making predictions (statistical sense) from data * Learning models from available training data, to make good predictions on unseen test data
• #28: Basic definition: Machine learning develops algorithms for making predictions (statistical sense) from data * Learning models from available training data, to make good predictions on unseen test data
• #29: Basic definition: Machine learning develops algorithms for making predictions (statistical sense) from data * Learning models from available training data, to make good predictions on unseen test data
• #30: Basic definition: Machine learning develops algorithms for making predictions (statistical sense) from data * Learning models from available training data, to make good predictions on unseen test data
• #31: Basic definition: Machine learning develops algorithms for making predictions (statistical sense) from data * Learning models from available training data, to make good predictions on unseen test data
• #32: Demo Power BI Desktop. Demos are available at //BI.
• #33: Basic definition: Machine learning develops algorithms for making predictions (statistical sense) from data * Learning models from available training data, to make good predictions on unseen test data