How Data Scientists Work
•Download as PPTX, PDF•
0 likes•60 views
Presentation to DBAs and SQL users about how data scientists work. Data scientists often don't know in advance what kind of data they need for building models. Therefore, it's hard to architect data bases around their (often transient) needs. We present a workflow and a few examples to illustrate the core process of training machine learning models.
Recommended
Recommended
More Related Content
Similar to How Data Scientists Work
Similar to How Data Scientists Work (20)
Recently uploaded
Recently uploaded (20)
How Data Scientists Work
- 1. © 2015 RentPath, LLC. All rights reserved. Data Science Péter Molnár <pmolnar@rentpath.com>
- 2. © 2015 RentPath, LLC. All rights reserved.© 2017 RentPath, LLC. All rights reserved. A design process 2 Purpose •Determine the purpose of the database •This helps prepare for the remaining steps. Organize Information •Find and organize the information required •Gather all of the types of information to record in the database Tables •Divide information items into major entities or subjects, such as Products or Orders. •Each subject then becomes a table. Columns •Decide what information needs to be stored in each table. Each item becomes a field, and is displayed as a column in the table. Primary Keys •The primary key is a column, or a set of columns, that is used to uniquely identify each row. Relationships •Look at each table and decide how the data in one table is related to the data in other tables. •Add fields to tables or create new tables to clarify the relationships, as necessary. Design Refinement •Analyze the design for errors. •Create tables and add a few records of sample data. •Check results. •Make adjustments to the design, as needed. Normalization •Apply the data normalization rules to see if tables are structured correctly. •Make adjustments to the table
- 3. © 2015 RentPath, LLC. All rights reserved.© 2017 RentPath, LLC. All rights reserved. Cross Industry Standard Process for Data Mining (CRISP-DM) https://en.wikipedia.org/wiki/Cross_Industry_Standard_Process_for_Data_Mining 3
- 4. © 2015 RentPath, LLC. All rights reserved.© 2017 RentPath, LLC. All rights reserved. Learning 4 • Supervised: • Un-supervised Learning
- 5. © 2015 RentPath, LLC. All rights reserved.© 2017 RentPath, LLC. All rights reserved. 5 Areas of Machine Learning Discrete Level Target Continuous Value Target Supervised Classification Regression Supervised Unsupervised Clustering Association Rules Unsupervised Disjoint Segmentation
- 6. © 2015 RentPath, LLC. All rights reserved.© 2017 RentPath, LLC. All rights reserved. Star Schema vs. Vector Space
- 7. © 2015 RentPath, LLC. All rights reserved.© 2017 RentPath, LLC. All rights reserved. Churn Prediction 7 • Who is at risk to leave? • Why are customers leaving? • Understanding factors that contribute to churn
- 8. © 2015 RentPath, LLC. All rights reserved.© 2017 RentPath, LLC. All rights reserved. Data Collection & Feature Selection/Engineering 8 Churn Stay
- 9. © 2015 RentPath, LLC. All rights reserved.© 2017 RentPath, LLC. All rights reserved. Training & Validation 9 yX Training ⎲⍙ ⎳⍦ y X Test ⎲⍙ ⎳⍦ Y Predicted Churn Stay Actual Churn True Positive We said customer churns and they left False Negative They left, we didn’t see it coming Stay False Positive We thought they would leave but they stayed True Negative We know our loyal customers
- 10. © 2015 RentPath, LLC. All rights reserved.© 2017 RentPath, LLC. All rights reserved. Evaluation 10 Predicted Churn Stay Actual Churn True Positive = 201 2.01% We said customer churns and they left False Negative = 99 0.99% They left, we didn’t see it coming Recall Churn = 0.67 Stay False Positive = 485 4.85 % We thought they would leave but they stayed True Negative = 9215 92.15% We know our loyal customers Recall Stay = 0.95 Precision Churn = 0.29 Precision Stay = 0.99 • Let’s assume 3% out of 10,000 customers churn • The model produces the above results
- 11. © 2015 RentPath, LLC. All rights reserved.© 2017 RentPath, LLC. All rights reserved. Why did they leave? • Many Machine Learning Algorithms are like a magic black box: we don’t know what’s going on inside • Sometimes it’s not about creating a the best model but trying to understand the factors that contribute to churn. 11 Decision Tree Logistic Regression
- 12. © 2015 RentPath, LLC. All rights reserved.© 2017 RentPath, LLC. All rights reserved. Recommendation Engine 12 Customize consumers experience Categorize properties for predictive modeling
- 13. © 2015 RentPath, LLC. All rights reserved. Which properties are similar? • Compare properties by – Price point – Amenities – Style – Location – Size – and many other qualities • Let Consumers define – Ratings of various properties by same consumer – Lead submission by same consumer to different properties 13 Pros • Assumption of similarity should hold for new inventory • Unbiased Cons • Biased • Our assumption about similarity might be wrong • Requires history data of consumer preferences • Fails on new inventory
- 14. © 2015 RentPath, LLC. All rights reserved.© 2017 RentPath, LLC. All rights reserved. ⎲⍙ ⎳⍦ Collaborative Filter 14 Items Users ⭐️ ⭐️⭐️ ⭐️ ⭐️⭐️ ⭐️⭐️ ⭐️ ⭐️⭐️⭐️ ⭐️ ⎲⍙ ⎳⍦ Item Attributes
- 15. © 2015 RentPath, LLC. All rights reserved.© 2017 RentPath, LLC. All rights reserved. Image Classification 15
- 16. © 2015 RentPath, LLC. All rights reserved.© 2017 RentPath, LLC. All rights reserved. Alex Net 16 https://papers.nips.cc/paper/4824-imagenet-classification-with-deep-convolutional-neural-networks.pdf
- 17. © 2015 RentPath, LLC. All rights reserved.© 2017 RentPath, LLC. All rights reserved. ImageNet http://image-net.org 17 Places205-AlexNet: CNN trained on 205 scene categories of Places Database (used in NIPS'14) with ~2.5 million images.
- 18. © 2015 RentPath, LLC. All rights reserved.© 2017 RentPath, LLC. All rights reserved. Summary 18 Discrete Level Target Continuous Value Target Supervised Classification Regression Supervised Unsupervised Clustering Association Rules Unsupervised Disjoint Segmentation ⎲ ⍙ ⎳ ⍦ y ⎲ ⍙ ⎳ ⍦ Y TP FN FP TN Péter Molnár <pmolnar@rentpath.com>
- 19. © 2015 RentPath, LLC. All rights reserved.© 2017 RentPath, LLC. All rights reserved. Summary 19 Discrete Level Target Continuous Value Target Supervised Classification Regression Supervised Unsupervised Clustering Association Rules Unsupervised Disjoint Segmentation ⎲ ⍙ ⎳ ⍦ y ⎲ ⍙ ⎳ ⍦ Y TP FN FP TN Péter Molnár <pmolnar@rentpath.com>
- 20. © 2015 RentPath, LLC. All rights reserved.© 2017 RentPath, LLC. All rights reserved. 20
Editor's Notes
- Data Science The talk aims to give an overview of common topics data science and how they may relate to database administration. Often the required data live on multiple database systems in structures that are different from the format that the analytics algorithm requires. On the large scale this can add pressure on the existing DB infrastructure, and slow down the data extraction process. Optimally, one would consider those requirement in the DB design process. However, many data science projects are for exploratory purposes and have a short lifespan. Dr. Péter Molnár is a data scientist at RentPath, LLC and faculty at the Institute of Insight at Georgia State University. As a academic and business professional, he is advancing and applying data science theories and tools in both public and private domains, including research in robotics, artificial intelligence, and machine learning.
- Data transformation: Star schema vs vector space