Operationalizing Data Science St. Louis Big Data IDEA
•Download as PPTX, PDF•
0 likes•135 views
The document provides an overview of the key steps for operationalizing data science projects: 1) Identify the business goal and refine it into a question that can be answered with data science. 2) Acquire and explore relevant data from internal and external sources. 3) Cleanse, shape, and enrich the data for modeling. 4) Create models and features, test them, and check with subject matter experts. 5) Evaluate models and deploy the best one with ongoing monitoring, optimization, and explanation of results.
Report
Share
Operationalizing Data Science St. Louis Big Data IDEA
- 1. • Please mute your phone and turn off your video. There are over eighty people who don’t want to see or hear you chewing. • If you have any suggestions for future topics that you would like this group to cover, please send them to Scott Shaw using Webex’s chat feature. • We will send out the presentation deck after the meeting. Look for an announcement in the Meetup link for this meeting. • If you have questions during the presentation, also send them to Scott Shaw using Webex’s chat feature. We will get to as many questions as we can. Before we begin…
- 2. Operationalizing Data Science Adam Doyle Daugherty Business Solutions April 1, 2020
- 3. Thanks for coming out everyone!
- 4. April Fools!
- 5. Operationalizing Data Science Adam Doyle Daugherty Business Solutions April 1, 2020
- 6. Begin with the end in mind. Pause in the middle to make sure that you can get to where you are going. • What is the business intention that you are trying to achieve? • Minimize Cost • Maximize Return • Minimize Risk • Realize Opportunity • Engage Stakeholders • POC vs Production ready and valued product
- 7. Identify your thesis. Goal vs. intention SMART goal Refine to question that can be answered with data science Data science – predict, explain, evaluate Decision science – combination of data science and data engineering
- 8. Acquire data. Third-Party Data Internal API Streaming General – Amount, Access, Quality, Labeled? Third Party o Assess Data Quality (Value Range, Adherence, Representative) o Data Format (Automatic vs hand-generated, Similar data from different partners are vastly different) o Governed (Use appropriate – avoid reidentification, TTL, Contractuals, Track access, renewals) Internal API (Data size limits, unreliability, costs) Streaming (CDC, Device Data, Standardized?)
- 9. Explore the data. Data Exploration Statistical Relationships and Correlations Profiling Textual – Word, Stop Words, Bigram, Trigram Clustering Check in with SME
- 10. Every block of stone has a statue inside it, and it is the task of the sculptor to discover it. Cleanse data. Data profiling Deduplication Outliers Filter Imputation Source Corrections Data shaping Sort Project Enrichment
- 11. Create the model and features. Type of Models (Supervised, Unsupervised, Reinforcement Learning, Neural Networks) Feature Engineering (Transformations and Aggregations) Encode Indicator Variables Binning/Bucketing Sparse Classes Interaction Features Extract Elements (eg. Time) Normalization Feature Selection Testing your features Testing your model Check in with SME
- 12. Check in with Business Does what you’ve created address the concerns of the business?
- 13. Batch vs. Real-time? Batch Training vs Real-time for - Training - Evaluation
- 14. Evaluate the model. Accuracy Precision Recall MSE Alignment to Business
- 15. Deploy the model. Automation Scaling SLAs Versioning Data Pipelines Ongoing Data Acquisition Ongoing Data Cleaning Ongoing Feature Encoding Integration in application
- 16. Monitor the model. Drift Degrading the model Predictions and their effects
- 17. Optimize the model. Feature Optimization Retraining Remodeling
- 18. Conclusion What does it mean to be done? Explanation as a Result
- 19. Questions?
- 20. • https://www.dataengineeringpodcast.com/ • https://dataengweekly.com/ • https://www.logicalclocks.com/blog/feature-store-the-missing-data- layer-in-ml-pipelines • https://www.imperva.com/blog/deployment-isnt-the-final-step- monitoring-machine-learning-models-in-production/ Links
Editor's Notes
- Welcome. Introduction.
- Welcome. Introduction.
- What is the business intention that you are trying to achieve? Minimize Cost Maximize Return Minimize Risk Realize Opportunity Engage Stakeholders POC vs Production ready and valued product
- Decision science SMART goal Goal vs. intention Refine to question Data science – predict, explain, evaluate
- General – Amount, Access, Quality, Labeled? Third Party o Assess Data Quality (Value Range, Adherence, Representative) o Data Format (Automatic vs hand-generated, Similar data from different partners are vastly different) o Governed (Use appropriate – avoid reidentification, TTL, Contractuals, Track access, renewals) Internal API (Data size limits, unreliability, costs) Streaming (CDC, Device Data, Standardized?)
- Data Exploration Statistical Relationships and Correlations Profiling Textual – Word, Stop Words, Bigram, Trigram Clustering Check in with SME
- Data profiling Deduplication Outliers Filter Imputation Source Corrections Data shaping Sort Project Enrichment
- Type of Models (Supervised, Unsupervised, Reinforcement Learning, Neural Networks) Feature Engineering (Transformations and Aggregations) Encode Indicator Variables Binning/Bucketing Sparse Classes Interaction Features Extract Elements (eg. Time) Normalization Feature Selection Testing your features Testing your model Check in with SME
- Check in with Business Does what you’ve created address the concerns of the business?
- Batch Training vs Real-time Training Batch Evaluation vs Real-time Evaluation
- Truth Matrix Mean Square Error Evaluation time
- Automation Scaling SLAs Versioning Data Pipelines Ongoing Data Acquisition Ongoing Data Cleaning Ongoing Feature Encoding Integration in application
- Drift Degrading the model Predictions and their effects
- Feature Optimization Retraining Remodeling
- What does it mean to be done? Explanation as a Result