Ml in a Day Workshop 5/1
Download as PPTX, PDF1 like183 views
This document provides an overview of machine learning and how it can be used by organizations. It describes machine learning as a way to generate new data through predictions and decisions by applying advanced mathematics and statistics to automate parts of the analysis. The document outlines how machine learning works by using algorithms to generate statistical models from data and then refine those models through an iterative process. It also provides examples of how machine learning is used in various industries like healthcare, finance, engineering, and services.
Ad
More Related Content
What's hot (20)
Similar to Ml in a Day Workshop 5/1 (20)
Ad
More from CCG (20)
![[Webinar] Top Power BI Updates You *Acutally* Need to Know](https://cdn.slidesharecdn.com/ss_thumbnails/powerbiwebinarmarch-200303213920-thumbnail.jpg?width=560&fit=bounds)
Ad
Recently uploaded (20)
Ml in a Day Workshop 5/1
- 1. Machine Learning 101 Advanced Analytics and Data Science
- 2. CCG Analytics Solutions & Services DATA MANAGEMENT Data & analytics consultants with a passion for helping clients overcome business challenges & increase performance by leveraging modern analytic solutions. BUSINESS ANALYTICS DATA STRATEGY
- 3. VOICES OF OUR CUSTOMERS “CCG to brought the expertise and the vision of to help us execute, to provide visibility to the data in a manner that we can use it faster.” - Gary Gray, Business Solutions Executive, Corsicana Mattress Company “The people we talked to know us. CCG wasn’t trying to fit us into a boilerplate template but prescribe a tailored solution. Their RapidRoadmap was the basis of our BI Strategy for the next two years.” - Kevin Davis, Sr. Director of BI, Kforce “Many times with CCG, we come to the table with questions or ideas and within a couple of days or weeks the team comes back with above and beyond what we actually asked for. They care.” - Chris Fitzpatrick, Vice President of Business Analytics & Strategy, vineyard vines “"I'm amazed at the talent at CCG, not just the skillset - they're really good people. We've already referred them once and will do so again!” - CIO, Ruth’s Chris Hospitality Group
- 4. Objectives By the end of this workshop, you should be able to: Describe what Machine Learning is and how it fits in to the analytic landscape Understand the difference between traditional and “advanced” analytics Describe what a statistical model is Understand a machine learning approach to statistical modeling The conceptual methodology behind the Machine Learning areas of classification, and clustering Describe some the most common tools for implementing data science and Machine Learning
- 5. AGENDA Why should anyone care about machine learning? What is Machine Learning? How does Machine Learning work? Ok but how does it really work? How can an organization use Machine Learning?
- 6. Machine Learning has tons of useful applications you already encounter or hear about every day. Analyzing Images Understanding Language Forming & Executing Strategy Personalized Recommendations Autonomous Decisions Predicting Asset Values How is Machine Learning used?
- 7. Machine Learning isn’t just applicable to high tech. There are suitable use cases present in most business sectors. Where is Machine Learning used? Healthcare • Claims Fraud • Real-time mortality risk for ICU patients • Response Adapted Radiotherapy • Predicting patient medication adherence • Translational/precision medicine Finance • Foreclosure/credit risk • Risk analysis • Fraud detection • Demand forecasting • Anti Money Laundering • Algorithmic trading Engineering • Resource allocation • Grid optimization • Robotics • Anomaly detection • Image recognition • Predictive maintenance Service • Single view of customer • Customer service analysis • Inventory planning • Social media analysis • Lead scoring • Marketing campaign evaluation
- 8. The concepts in Machine Learning are not new, however its use has increased How has Machine Learning Evolved? https://www.quantinsti.com/blog/machine-learning-basics
- 9. Even though the concepts are decades old, machine learning has only become more feasible in recent years. Why Machine Learning Now? Flood of data and decreasing costs of storage Increasing computational power Increased attention from researchers Growth of open source technologies Support from industries
- 10. AGENDA Why should anyone care about machine learning? What is Machine Learning? How does Machine Learning work? Ok but how does it really work? How can an organization use Machine Learning?
- 11. Advanced Analytics deal with making “best guesses” faster, better, and more consistent than relying on human SMEs. Provide insights to existing data using: • Raw data points • Summaries of data • Calculations across existing data fields • KPIs The data reported are historical or current facts. “Just give me a report so I can make a decision.” Generally requires the application of basic mathematics or arithmetic. Generate new data, including: • Predicted future values • Best guesses of missing values • Suggested next steps • Categorizations The data generated are “best guesses” and contain some uncertainty. “Predict what might happen so I can make the best possible decision.” Requires the application of advanced mathematics, statistics and computing principles. Traditional Analytics Advanced Analytics Traditional vs. Advanced Analytics
- 12. Advanced Analytics (“AA”) enable predictive and prescriptive uses of data by applying sophisticated math and statistics to automate parts of the analysis. What is Advanced Analytics? Traditional analytics focuses on understanding and explaining the data that has been collected. AA focuses on generating new data in the form of predictions or decisions, and going the extra step to automate decision-making when possible.
- 13. Machine Learning contains the tools that allow the application and execution of Advanced Analytics. What is Machine Learning? Machine learning is . . . Computationally intensive Uses advanced mathematics Designed to self-optimize
- 14. AGENDA Why should anyone care about machine learning? What is Machine Learning? How does Machine Learning work? Ok but how does it really work? How can an organization use Machine Learning?
- 15. A model is a repeatable, data-driven approach to making a best guess. It does this by formalizing mathematical relationships between data in the form of either: – Rules (e.g. predict applicants will default on a loan if Credit Score < 700 and Debt to Income Ratio > 30%) – Or an equation (e.g. predict Home Price = 100*Square Footage + 2*Average Income in the Area) NOTE that this is not the same as a DATA model. These are different things: Machine Learning works by using algorithms to generate “models.” How does Machine Learning work? Data Model Statistical Model
- 16. In the past we’ve told computers how to use data to a answer our questions. Data Prior month sales: $4MM 2 months prior: $3MM 3 months prior: $2MM Program / Model This month sales = (prior month + 2 months prior + 3 months prior) / 3 Answer This month’s sales = $3MM? What’s a model?
- 17. Answer Last month’s sales: $2MM Data Prior month sales: $4MM 2 months prior: $3MM 3 months prior: $1MM Answer Last month’s sales: $2MM Data Prior month sales: $4MM 2 months prior: $3MM 3 months prior: $1MM Answer Last month’s sales: $2MM Data Prior month sales: $4MM 2 months prior: $3MM 3 months prior: $1MM Answer Last month’s sales: $2MM Data Prior month sales: $4MM 2 months prior: $3MM 3 months prior: $1MM But we’ve found that if we give the machine historic facts, we can let it find the right program / model to plug in for future answers Answer Last month’s sales: $2MM Data Prior month sales: $4MM 2 months prior: $3MM 3 months prior: $2MM Program / Model This month’s sales = 1/8 * Prior month + 1/3 * 2 months prior + 1/4 * 3 months prior What’s a model?
- 18. Answer Last month’s sales: $2MM Data Prior month sales: $4MM 2 months prior: $3MM 3 months prior: $1MM Answer Last month’s sales: $2MM Data Prior month sales: $4MM 2 months prior: $3MM 3 months prior: $1MM Answer Last month’s sales: $2MM Data Prior month sales: $4MM 2 months prior: $3MM 3 months prior: $1MM Answer Last month’s sales: $2MM Data Prior month sales: $4MM 2 months prior: $3MM 3 months prior: $1MM Once we have our machine-defined program, we can use it with new data to make better predictions Answer Last month’s sales: $2MM Data Prior month sales: $4MM 2 months prior: $3MM 3 months prior: $2MM Program / Model This month’s sales = 1/8 * Prior month + 1/3 * 2 months prior + 1/4 * 3 months prior New Data Prior month sales: $8MM 2 months prior: $6MM 3 months prior: $8MM Answer This month’s sales = $5MM What’s a model?
- 19. A defined set of steps for solving a problem Often involves repeating steps May or may not have an ending condition – The problem is solved to our satisfaction • For example – stop when the last 4 iterations have been 95% accurate or better – The problem hasn’t been solved but we don’t seem to be getting any closer to solving it • For example – stop if the last 10 iterations have not seen any improvement in accuracy – The process has run for a long time • For example – stop after the program has run for 12 hours, regardless of whether progress is still being made The word algorithm gets used a lot, but it isn’t always defined. What is an algorithm? Algorithm to sum a row of numbers
- 20. Collect the data and randomly create initial decision rules. Design a method for evaluating how good your hypothesis is, and test whether it applies generally. Update your hypothesis in a way that marginally improves the performance of your decision rules. Continue this process until the hypothesis either you are satisfied with the results, or your hypothesis can’t improve anymore with the data available. Almost all machine learning algorithms follow the same general pattern. They work best when they have large amounts of data to test the hypothesis. Create a hypothesis Evaluate the hypothesis Adjust the hypothesis Repeat until convergence What is an algorithm?
- 21. Machine learning is making models better, faster, and stronger. Technical advances in the past several years have enabled machines to conduct rapid trial and error with massive amounts of data, resulting in better models that can be achieved quicker. Machine Learning • Machine learning algorithms are designed to make the most accurate predictions possible. • Algorithms are defined to find solutions via rapid iterations, where slight improvements are made with each new piece of information. • Machines can consider thousands of variables to design the optimal model • Models can be tested in hours or minutes • Models can be designed to self-adapt in real time to changes in the environment • Advanced Machine Learning techniques can solve complex problems like interpreting visual or audio input using unstructured data Statistical Approaches to Modeling • Traditional statistical models are designed for inference about the relationships between variables. • Simplistic models are often solved mathematically, but improvements require guesswork • Human experts must select a reasonable amount of input data for hypothesis testing when designing their models • Humans may only be able to evaluate a handful of models over the course of weeks or months • Models are reviewed periodically (often annually or even less frequently) for adjustments • Traditional modeling techniques require structured data and are too simplistic to emulate human intuition How is machine learning changing the game?
- 22. AGENDA Why should anyone care about machine learning? What is Machine Learning? How does Machine Learning work? Ok but how does it really work? How can an organization use Machine Learning?
- 23. There are two main families of algorithms to choose from. Supervised Learning Unsupervised Learning There aren’t necessarily “right answers,” we just want to get a better understanding of our data. We know the “right answers” for some of the scenarios. – We may have history we can look back on – We may be hoping to replicate human decision making
- 24. Supervised or Unsupervised? Predict our profits next quarter. Supervised Identify the number written on a check. Group our customers into segments. Supervised Unsupervised Predict a user’s rating for a given product. Supervised Find the most important variables in a dataset. Unsupervised Identify credit card transactions that are out of the ordinary. Unsupervised
- 25. Now let’s walk through two of the most popular machine learning approaches and discuss how the algorithms are applied. How does an algorithm really work for businesses? Classification Clustering
- 26. Use classification when you want to guess a non-numeric value, like a yes/no answer. We will take a decision tree approach. Everyone will repay their loan. Create a hypothesis 20 outstanding loans
- 27. Use classification when you want to guess a non-numeric value, like a yes/no answer. We will take a decision tree approach. Calculate accuracy as the % of predictions that are correct based on your current set of rules. Evaluate the hypothesis 20 outstanding loans 12 repaid, 8 defaulted Accuracy = 12/20 = 60%
- 28. Use classification when you want to guess a non-numeric value, like a yes/no answer. We will take a decision tree approach. Find the next branch by looking for the data split that would have the biggest impact on the purity of each node. There are several ways to do this mathematically (Gini Index, Information Gain, Chi- Square). Adjust the hypothesis 20 outstanding loans20 outstanding loans 20 outstanding loans Credit Score > 700Credit Score < 700 Income > 60kIncome < 60k DTI > 40%DTI < 40% 80%73%70%50%71%53% 59% weighted 60% weighted 75% weighted
- 29. Use classification when you want to guess a non-numeric value, like a yes/no answer. We will take a decision tree approach. Repeat the process for each of your new “leaf” nodes. Stop when you reach an acceptable level of accuracy, or when your accuracy begins getting worse with independent data. Repeat until convergence 20 outstanding loans DTI > 40%DTI < 40% Credit Score > 700Credit Score < 700Income > $60kIncome < $60k 100%50% 100%100% 80% weighted
- 30. Classification is used for lots of problems that copy human intuition. Think about how you classify information to identify these images! These use cases are obviously more complex than our simple decision tree, but with more advanced approaches like convolutional neural networks these pictures can definitely be classified by a machine.
- 31. Use clustering when there’s no “correct” classification, but you still want to assign individuals to groups. This algorithm is called k-means clustering. If I want 3 groups, where would I as a human draw the boundaries?
- 32. Use clustering when there’s no “correct” classification, but you still want to assign individuals to groups. This algorithm is called k-means clustering. I can segment my customers by assigning them to 3 groups. We’ll set down 3 random “anchors” and assign each customer to its closest anchor. Create a hypothesis
- 33. Use clustering when there’s no “correct” classification, but you still want to assign individuals to groups. This algorithm is called k-means clustering. Find the distance between each customer and each of these “anchors”. Take note of which customers are actually closest to a different anchor than the one they’re assigned to. Evaluate the hypothesis
- 34. Use clustering when there’s no “correct” classification, but you still want to assign individuals to groups. This algorithm is called k-means clustering. Reassign each customer to the group corresponding to the anchor they’re closest to, and move the anchors to the middle of their new group. Adjust the hypothesis
- 35. Use clustering when there’s no “correct” classification, but you still want to assign individuals to groups. This algorithm is called k-means clustering. Repeat until convergence Keep moving the anchors and re-assigning customers until the anchors stop moving.
- 36. This is just the tip of the iceberg. There are several algorithms available for various types of problems.
- 37. AGENDA Why should anyone care about machine learning? What is Machine Learning? How does Machine Learning work? Ok but how does it really work? How can an organization use Machine Learning?
- 38. Delivering analytics with Machine Learning requires alignment across people, process, technology, and data. Engaging with Machine Learning Image inspired by Microsoft People Process Technology Data Guide Support Enable
- 39. Data scientists combine broad skills to integrate data, build models, and drive business value. People Process Technology Data
- 40. Let’s look at the Microsoft TDSP to see how data scientists spend their time. People Process Technology Data
- 41. Traditional Analytics The outputs of the process can be used in traditional analytics, analyzed directly, or fed into automated decision-making. Store and access data. Filter and aggregate it. Visualize it. Show it to the business so they can take action. Machine Learning Filter and aggregate it. 1 𝑁 𝑛=1 𝑁 𝑥 Create a model. Generate new data (predictions, etc.). The new data can be stored with the rest of the data for use in analytics. Or it can be visualized directly to gain insights. Or it can automate decisions or actions, allowing better processes to run faster and 24/7. People Process Technology Data
- 42. The sources of data for use in data science can be broad. People Process Technology Data Data Warehouses •Curated & Governed data •Big data •Cloud or on-prem Data Lakes •Unstructured & Semi-structured data •Streaming data •Partially curated Externally Procured Data •May be purchased from 3rd party providers •May be scraped from the web •May require designing research experiments Data scientists typically have the programming and data integration skills to use data from anywhere it can be found.
- 43. The Microsoft technology stack provides a holistic solution to your Machine Learning needs. People Process Technology Data
- 44. CCG’s Rapid Insight Solution Actionable Backlog – Of use cases ripe for predictive analytics to transform your business Detailed Readouts – The materials we leave behind will include extensive analysis of our methodology, findings, and recommendations Ownership of the Model – Just because the project ends doesn’t mean the model stops working. Unlike other managed service providers, what we produce on your behalf is yours to keep Identify Use Cases – By holding a workshop with process SMEs to identify opportunities to supercharge the business Summarize the Findings – So you can understand the model’s outputs and begin taking action on what we’ve learned Develop a Prototype Model – To generate forecasts, classifications, or exploratory analysis for one of your use cases using an industry-standard tool like Azure Machine Learning Studio or Databricks Week 1 Weeks 2-5 Week 6
- 45. Fully Operational Production Model – Available at all times, in production – Batch & API integrations Model Supervision – Model is monitored for ongoing usability – Performance dashboard – Guaranteed accuracy SLAs Model Retraining & Support – Scheduled & triggered model re-tuning or re-training – Add new data features over time Model as a Service Solution Set up model as a web service Visualize model performance in a dashboard Maintain and enhance model
- 46. THANK YOU! What questions do you have?
- 47. Microsoft offers pre-built APIs through Cognitive Services that can expedite the deployment of AI capabilities. People Process Technology Data
- 48. VISUAL DRAG-AND-DROP Azure Machine Learning Studio
- 49. What is Azure Databricks? A fast, easy and collaborative Apache® Spark™ based analytics platform optimized for Azure Best of Databricks Best of Microsoft Designed in collaboration with the founders of Apache Spark One-click set up; streamlined workflows Interactive workspace that enables collaboration between data scientists, data engineers, and business analysts. Native integration with Azure services (Power BI, SQL DW, Cosmos DB, Blob Storage) Enterprise-grade Azure security (Active Directory integration, compliance, enterprise -grade SLAs)
- 50. Azure Databricks key audiences & benefits Unified analytics platform Integrated workspace Easy data exploration Collaborative experience Interactive dashboards Faster insights • Best of Spark & serverless • Databricks managed Spark Improved ETL performance • Zero management clusters, serverless Easy to schedule jobs Automated workflows Enhanced monitoring & troubleshooting • Automated alerts & easy access to logs Zero Management Spark Cluster democratization (serverless) Fast, collaborative analytics platform accelerating time to market No dev-ops required Enterprise grade security • Encryption • End-to-end auditing • Role-based control • Compliance Data scientist Data engineer CDO, VP of analytics Provided by Microsoft and Databricks under NDA
- 51. Machine Learning combines statistics and computer science to achieve advanced analytics outcomes. What is Machine Learning? Computational Power Statistics Predictive & Prescriptive Analytics Faster More Accurate
Editor's Notes
- #4: Quote from PGI taken from website.
- #5: We’ll be discussing a lot of material this morning as well as in this workshop in general. By the end of this workshop we are hoping that you’ll be able to: Describe Machine Learning and how it fits in to the greater analytic landscape. Have a solid understanding about the differences between traditional and advanced analytics Be able to describe what a statistical model is Understand how Machine Learning is used in statistical modeling Classification and clustering and the concepts used in applying those models And finally, to describe some of the most common tools used in applying Machine Learning.
- #6: Why should anyone care about Machine learning?
- #7: We should care because Machine Learning applications are all around us: Predicting property values – Zillow Image analysis – The postal service uses optical character recognition (OCR) to identify and sort zipcodes. Facial recognition is now seeing wider use at airport screening checkpoints and airline ticket counters. Understanding language – Cortana and SIRI are used in smartphone technology Recommendation engines are used by Amazon, Netflix and other companies. “based on your “preferences” we think you’ll like . . .” Forming and executing strategy – Deep blue played chess again world chess champion Garry Kasparov and IBMs Watson competed in Jeopardy and destroyed its human opponents Autonomous decisions – Tesla produces self-driving cars.
- #8: Although there are a seemingly infinite number of machine learning applications in high-tech, there are many use cases are present in other sectors as well. In healthcare, ML is used to detect claims fraud. I used to work for a health insurance provider and we had machine learning algorithms running as claims were electronically submitted in order to help detect fraud. There are methods that can help customize treatment plans for patients with cancer and other chronic diseases, and by using vital signs, laboratory and other data from a patients electronic medical record, there are methods that can perform real-time monitoring and risk assessment of patient mortality risk. In finance, it is used to determine creditworthiness and default risk and it does this in real-time. When a credit card or loan application is submitted over the internet, the application is evaluated according to the algorithm and an immediate decision is often rendered. There are also wide ranging applications in the engineering and service industries. For example, predictive maintenance is on the increase in the airline industry. The electronic components in transport category aircraft all operate within certain parameters. When those parameters are exceeded even for a millisecond, a log file and a data point are created. Mechanics and aircraft engineers can analyze these to determine when a part may fail completely and thus, replace it proactively and improve operational efficiency. I’m sure we’ve all at some point been delayed because of a maintenance problem on an aircraft. Predictive maintenance aims to help reduce that.
- #9: Where did machine learning come from and how has it evolved over the years. Well, these concepts aren’t new. In 1950, the “Turing Test” was developed to see if a human could tell whether they were talking to another human or a machine. The perceptron, which we know as a neural network was developed in 1957. In 1997, Deep Blue beat world chess champion Garry Kasparov in a six match series. Interestingly, in 1996, Kasparov beat Deep blue 4 games to 2. So if these concepts aren’t new, why has their utilization increased so much?
- #10: Flood of data and decreasing costs of storage - With the advent of new technologies and data compression algorithms, storage, which was historically very costly, is now more economical than ever. Data is everywhere now. Email and the internet make data easy to find and retrieve and also easy to transfer. Large datasets which once would have required physical storage disks to transport, can now be sent online very quickly via an FTP server. Increasing computational power – It seems like every week or month, a new faster processor is on the market. Multi-core and cluster computing architectures make execution times orders of magnitude faster than they were just a few years ago. Increased attention from researchers – Machine Learning methods have reduced the time required for heavy computational research from months or weeks, to days or hours. With this major time savings also comes major cost savings. Growth of open source technologies – With open source computing technologies such as Python, R, Apache Spark, and others, the need for heavy capital investment in hardware and software has been substantially reduced which makes running machine learning algorithms easier and more cost effective. Support from industries – Machine learning has been transformative for industry, the IDC (international data corporation) forecasts that spending and investment on machine learning and artificial intelligence will increase from $12 billion in 2017 to $57.6 billion by 2021.
- #11: What is Machine Learning?
- #12: In order to understand what machine learning is, we need to understand what traditional and advanced analytics are. Historically speaking, analytics has generally been raw data and summaries of raw data within an organization. This data could be found in a number of places. Perhaps formal databases, excel spreadsheets, post-it notes on someone's desk, etc. These could be calculations of sales, inventory levels, KPI’s etc. The calculations were descriptive in nature, requiring the application of basic math or arithmetic. Business decisions were often made on “gut feel”. Advanced Analytics is a term that has become more prevalent over the last few years and uses an organizations existing data to generate new additional data which are then combined in order to make predictions on “what might happen” or “what could happen in the future”. And although advanced analytics usually uses advanced math and statistics, there is always an element of uncertainty.
- #13: When we look at the continuum of different types of analytics, there are 4 distinct types. At it’s most simplistic, there are descriptive and diagnostic analytics. “What happened and why?” These are retrospective in nature and make up what we call traditional analytics, using an organizations existing information. As we move forward and up along the continuum, analytics are used to answer the questions, “what will happen” and “how can we make it happen”. Predictive and Prescriptive analytics allow us to optimize business decisions by automating the analysis and also the decision-making process when feasible. One note about the difference between Predictive and Prescriptive analytics. Although predictive analytics has an array of techniques that can be used in practice, the definition of prescriptive analytics doesn’t generally encompass the technical approaches by which it’s done. Generally speaking, the output from prescriptive analytics are decisions, not predictions.
- #14: Ok. So what about Machine Learning? Machine Learning is “the how”, it is how Advanced Analytics get done. As we can see on the continuum, Predictive and prescriptive analytics are the most difficult to execute, however they also provide the most value to the business. In terms of difficulty, these types of analytics are in line with the attributes of machine learning. Machine learning is computationally intensive, uses upper-level mathematics such as statistics, linear algebra and calculus, as well as advanced computing principles. It leverages computing power in order to automate the delivery of analytic output. Whether that is in the form of reports, or dashboards, etc. and as these algorithms run, they are also able to “learn” so-to-speak and optimize themselves in order to improve their predictive accuracy.
- #16: We use those algorithms, or set of steps, to design and generate models. A model uses the data in a dataset to make a prediction or best-guess about what might happen, based on the terms, or “parameters” of that model. For the most part, this happens in one of two ways. One, it creates a mathematical relationship within the data based on a set of rules (if this then that, or else that, etc.) Or, it uses an equation based on pre-specified terms in the model, or the algorithms can be set to select the most relevant terms automatically (called automatic model selection). The term “model” can also mean multiple things. We need to differentiate our statistical model from a data model (which is used in creating and formalizing relational databases) as they are not the same.
- #20: What do we mean by “algorithm”? We say it all the time, but never really talk about what it is. An algorithm is a defined set of steps for solving a problem. It often involves repeating steps or iterations, and the output of one iteration is typically part of the input for the next iteration. An algorithm may or may not have a point where it’s “finished” so to speak. Most of the time, the data scientist or statistician will specify the criteria at which the process should stop. Although the algorithm is well-defined set of steps, the algorithm is then translated into the form of a program which is then executed over a data set to produce the analytic output.
- #21: Most machine learning algorithms work in the same way and require an initial guess or hypothesis about what you might expect to find. We then evaluate that hypothesis by selecting an appropriate algorithm and model to answer that question. This is often challenging and multiple models may need to be created in order to compare them against each other. Based on the results of the initial model runs, you may then adjust your hypothesis in such a way as to improve the performance of your model. Finally, this process is repeated until you are either satisfied with the performance of your model, or no further improvements can be achieved after multiple attempts. So, CREATE – EVALUATE – ADJUST - REPEAT
- #22: There is a lot of information on this slide, but the key point here really is that a machine learning approach to modeling makes things much more achievable. Traditional statistical models are designed to examine whether significant relationships exist between the variables in the model, whereas machine learning is designed to make highly accurate predictions and help inform decision making. The nature of machine learning methods lends itself to automation and while traditional statistical models can be scheduled to run iteratively, improvements in the model usually require manual intervention by a data scientist or statistician. When testing hypotheses using traditional modeling, we have to manually select variables to be tested. Machine learning algorithms can evaluate thousands of variables during the modeling process. In the healthcare field, this is particularly advantageous when comparing individual genomic profiles. With around 24,000 genes in the human genome, it’s easy to see why an automated approach to analysis would be desirable.
- #23: Why should anyone care about Machine learning?
- #24: In machine learning, there are two main families of algorithms from which to choose. There is Supervised learning which is used when we already know the outcome for some of the scenarios and we want to train a model to replicate that outcome going forward. And Unsupervised learning which is used when we don’t necessarily know the outcome that we’re looking for, but we need a better understanding of the relationships that may be hidden in our data.
- #25: Quiz. Engage audience.
- #26: So how can you apply these in a business context? Let’s take a closer look at how two of the most popular machine learning algorithms are used in practice.
- #27: So our initial hypothesis is that “everyone will repay their loans”. Empirically, we know this not to be true, however it’s a perfectly acceptable starting hypothesis. Next let’s evaluate.
- #28: Based on our starting hypothesis of a 100% repayment we see that we are correct only 60% of the time.
- #29: Now that we’ve determined an initial accuracy of 60%, we then look to see how we could adjust our hypothesis to improve it. We will add branches to our tree using additional features in the dataset that will help improve the identification of those people who are likely to repay or default on their loans. After looking at income, Debt-to-Income ratio, and credit score, we find that Debt-to-income ratio provides the best information to help us discriminate between those that pay and those that default.
- #30: This process is repeated until an acceptable accuracy rate is achieved. In practice, thousands of these trees can be created quickly in order to find the best accuracy rate.
- #32: Just by looking at the raw data plotted out, we may think that we can identify logical groups based on the position of the datapoints. There would likely be multiple ways to group this data just using the human eye. But we’d like to use a machine learning approach that selects the optimal grouping of the data.
- #33: Our hypothesis is that we have three distinct groups within the data. Determining the number of groups in a clustering analysis is always an interesting challenge. There are a number of statistical methods which can help select the “right” number of groups. However, in practice, that generally becomes a business decision based on an educated guess.
- #34: Next we calculate the distance between each datapoint and each one of the anchors that we laid down. We’ll pay specific attention to the customers that are actually closer to a different anchor than the one they’re assigned to.
- #35: We then reassign each customer to the group corresponding to the anchor they’re closest to. We then move the anchors themselves to the middle of their new group.
- #36: We continue this process until the anchors stop moving. So then the question becomes, when will they stop moving? The anchors will stop moving when the distance between the anchor and all of its assigned customers has been minimized. The algorithm is then said to have “converged”.
- #37: There are a large number of machine learning algorithms, from the simplistic such as linear regression, to the complex such as deep learning methods and various neural network architectures. What you see here is just a sample of what’s available and many of these have minor variants that help improve the accuracy of the algorithm under certain conditions and in certain contexts.
- #38: Why should anyone care about Machine learning?
- #39: Engaging with Machine learning within an organization requires alignment across people, process, technology and data. The technology enables the people and supports the process which guides the people. Let’s take a look at each one of these.
- #40: People – many machine learning practitioners are data scientists. What exactly is a data scientist? A data scientist is someone with a broad skillset including mathematics and statistics, computer science and business acumen. He or she usually has substantial expertise in one of those areas and is proficient in one or more programming languages. Additionally, the ability to put structure and definition around sometimes vague or unclear business questions is critical. The practice of machine learning combines math and statistics, with computing principles and information technology. When you layer in the business and domain knowledge, you get data science. There are really two different types of data scientists. There is the data professional who has formal training in math and statistics and how to build statistical models, and then there is what we call the citizen data scientist who doesn’t have that formal training, but does have some experience and is proficient in building and deploying statistical models. Although citizen data scientists are do have experience, we don’t want to discount the importance of statistical rigor. Finding someone who has all these skills is really difficult. Therefore, it’s important to keep skillsets in mind when trying to build a data science team.
- #41: The data science process always begins with a business question or questions and it is essential to fully understand the context and substance of what is being asked. We need to remain focused on what is going to be actionable and valuable to the business. We need to understand how the model’s outputs will be used and we need to manage expectation so that stakeholders understand that data science is exploratory in nature and explanatory relationships may not exist in the data. Data scientists probably spend 40% of their time on data wrangling and another 40% in modeling and feature engineering. Once a model has been created through iterative development and testing, it should be deployed and made available to business users. There are several ways to do this including storing the output data in a database, deploying to a web service or REST API, or building the model directly into an application.
- #42: The analytic output from the data science process can be stored with the rest of the organizations data and subsequently, used in other modeling processes, or it can be used to create dashboards and deployed for business users and stakeholders to make decisions.
- #43: Without data, we don’t have analytics. The sources of data are abundant today, in both structured and unstructured forms. Data can be contained in traditional data warehouses or imported and stored in a data lake and then combined with other data that may be unstructured. One very important skill a data scientist needs is to be a data-wrangler (or data-munger), which is the ability to take data in both structured and unstructured data and form it into a cohesive clean dataset that can be analyzed.
- #44: So we’ve talked about the people we need to do machine learning, the process by which we do it, that types of data that we use. So what about the tools and technology. . . The Microsoft technology stack contains a very robust toolkit that addresses a wide range of use cases. Microsoft Cognitive services offers pre-built APIs that can speed up the deployment of AI capabilities. Machine learning frameworks such as Pytorch, TensorFlow and Keras are available to data scientists to design and build their own machine learning models. Command-line tools and UI tools such as Azure Databricks and Azure ML Studio enable both the power-user as well as the citizen data scientist. These tools support different processing architectures including GPUs and FPGA (field-programmable gate arrays) to provide major speed advantages to machine learning models. Finally, they can be deployed to hardware on-prem, in the cloud, or at the edge of IoT.
- #48: Microsoft Cognitive services contain pretrained models to accelerate solution development in areas such as Natural Language Processing, Speech recognition, collaborative filtering and recommendation engines, as well as image recognition models.
- #49: This is an example of the Azure ML Studio user interface. As you can see modules are all drag-and-drop. This is a very user friendly tool which integrates seamlessly with other Azure products. Models can be built, deployed and refreshed with just a few clicks.
- #50: What is Azure Databricks? Azure Databricks really helps to enhance the productivity of the data scientist. New clustering and computing environments can be created quickly in a collaborative, unified environment. There is native integration with Power BI, as well as the rest of the Azure platform. For those of you that may be familiar with apache spark, databricks is designed based on the Spark cluster-computing analytics engine.
- #51: There are a number of advantages Azure Databricks. For the data scientist, databricks has an integrated workspace where you can easily switch between R, python, and Scala. For the data engineer, its offers improved ETL performance and automated workflows. And as an analytics executive, you can sleep well at night knowing that Databricks offers role-based security, has full auditing capabilities and if you’re in the healthcare industry, is HIPAA complaint.
- #52: But why? Why is Machine learning so valuable to an organization? There are several reasons. Machine learning allows organizations to obtain analytics faster, which subsequently allows for faster business decision making. As data is collected and analyzed in real-time, machine learning models can re-execute quickly providing new predictions on demand, or on a predefined schedule. They are more accurate. Machine learning methods, when properly trained, will provide more accurate predictions over the course of time. The precise mechanism by which this occurs is highly technical, however, generally speaking, the algorithms are designed in such as way to “learn” from their own errors. By analyzing the difference between their predictions and what actually occurs, the algorithms are self-optimizing, meaning that they are adjusted in such a way so that the error rate gets smaller and smaller with each subsequent iteration.