D7 MarkPlus - Machine Learning Algorithm.pdf
- 1. Machine Learning Algorithm Novan Parmonangan Simanjuntak Head of Machine Learning and AI Strategy novanps novan.p.simanjuntak@glair.ai glair glair.ai hi@glair.ai Contact Us
- 2. 01 02 03 04 05 glair.ai Intro to Machine Learning Machine Learning Workflow 01 02 OUTLINE 2
- 4. AI ML A program that can sense, reason, act, and adapt Algorithms learns from data Artificial neural networks (inspired by brain) learn from data ARTIFICIAL INTELLIGENCE MACHINE LEARNING DL AI, ML & DL DEEP LEARNING 4 towardsdatascience.com
- 6. Why ML: Example Output/Label: Do I want to go to outside on the weekend given data? (Answer: Yes or No) Input: ● Weather: Rainy, Sunny ● Distance to destination (km) ● Invited by friends: yes, no Rules: 1. Initially, score = 0 2. If the weather is rainy then (score + 10), else if the weather is sunny then (score - 10) 3. score + distance 4. If invited by friends then (score -10), else if not invited by friends then (score + 10) 5. If score < 10 then Yes, else No No Weather Distance Invited by Friends Score Go on Weekend? (Answers) 1 Rainy 8 Yes 10+8-10 = 8 Yes 2 Sunny 20 No -10+20+10 = 20 No 3 Rainy 5 No 10+5+10 = 25 No Data: 6
- 7. Why ML: Example Let’s model it using AI (Model: Artificial Neural Network) 7
- 8. Why ML: Example Let’s model it using Artificial Neural Network 8 >=10 Example: ● Input: weather = Rainy, Distance = 8, Invited by Friends = Yes ● Score = 1 * 10 + 0 * -10 + 8 * 1 + 1 * -10 + 0 * 10 = 8 ● Score < 10, so the Output is Yes
- 9. Why ML?: A New Paradigm 9 Traditional Programming Input Rules Output Machine Learning Input Output Rules
- 10. Why ML? No Weather Distance (km) Invited by Friends Go on Weekend? (Answers) 1 Rainy 8 Yes Yes 2 Sunny 20 No No 3 Rainy 5 No No ML: Find the Rules 10
- 11. Optimization : Finding Rules 11 ● Determine Loss Function ● Update Rules/Weight to minimize loss
- 12. Why ML? If you know the Rules Do Not Use ML If not Use ML 12
- 13. Why DL? 13
- 14. 14
- 15. AI Approach Learning Type ● Supervised Learning ● Unsupervised Learning ● Reinforcement Learning 15
- 16. ● Input and Output is Given ● Finding a rules/function f that maps a set of points X (input/predictor) to a set of labels Y (output), based on given data (xi , yi ). ● Categorized to: ○ Classification (categorical labels) ■ Credit Scoring ■ Fraud Detection ■ Recommendation System ■ Object Recognition ■ Spam Filtering ○ Regression (numerical labels) ■ Home Price Prediction ■ Stock Market Prediction ■ Demand Forecasting Supervised Learning 16
- 17. Unsupervised Learning ● Only Input is given ● Find hidden patterns or underlying structure in the given data Document Clustering 17 search.carrotsearch.com/pertamina
- 20. 20 youtube.com
- 21. 21 youtube.com
- 22. glair.ai Reinforcement Learning in Action 22
- 24. glair.ai Machine Learning Workflow Business Problem ML Problem Framing Data Ingestion 24 24
- 25. • Transportation of data from varied sources to a storage where it can be accessed, used, and analyzed by an organization. • The destination is typically a data warehouse, data mart, or database. GDP LABS CONFIDENTIAL Data Ingestion 01 25
- 26. glair.ai Machine Learning Workflow Business Problem ML Problem Framing Data Preparation Data Ingestion 26
- 27. ● Make it easier for ML algorithm to interpret the data. ● Categories: ○ Data cleaning: ■ Garbage In Garbage Out ■ identifying incomplete, incorrect, inaccurate or irrelevant parts of the data and then replacing, modifying, or deleting the dirty data ○ Data encoding, normalization, resampling ○ Data Splitting (Training, Validation, Test) GDP LABS CONFIDENTIAL Data Preparation 02 27
- 28. glair.ai Machine Learning Workflow Business Problem ML Problem Framing Data Preparation EDA Data Ingestion 28
- 29. ● Exploratory Data Analysis Analyze data to summarize main characteristics, often using statistical techniques or data visualization. ● Objectives: ○ Suggest hypotheses ○ Assess assumptions ○ Support the selection of appropriate tools, techniques, and features GDP LABS CONFIDENTIAL Exploratory Data Analysis (EDA) 03 29
- 30. glair.ai Machine Learning Workflow Business Problem ML Problem Framing Data Preparation EDA Feature Engineering Data Ingestion 30
- 31. ● Feature engineering can use domain knowledge to extract new features from raw data ● Objectives: ○ Improving the performance of machine learning models. ● Examples: ○ Grouping operations over a window (e.g, average) ○ Binning ○ Log transformation GDP LABS CONFIDENTIAL Feature Engineering 04 31
- 32. glair.ai Machine Learning Workflow Business Problem ML Problem Framing Data Preparation EDA Feature Engineering Model Training and Tuning Data Ingestion 32
- 33. ● Choose which model to try ● Simpler model is better ● Iteratively using a more complex model and features if needed ● Use a baseline model ● Maximize model performance using hyperparameter tuning ● Consider constraints: ○ cost, explainability, and speed GDP LABS CONFIDENTIAL Model Training and Tuning 05 33
- 34. glair.ai Machine Learning Workflow Business Problem ML Problem Framing Data Preparation EDA Feature Engineering Model Training and Tuning Model Evaluation Data Ingestion 34
- 35. ● Evaluate the model with the test data ● Pick a suitable metrics for the problem. ○ There must be a Business and ML metrics for the problem ● If the business metrics achieved then continue with model deployment, if not reiterate the model creation with data augmentation (adding more data) or feature augmentation (adding other features) GDP LABS CONFIDENTIAL Model Evaluation 06 35
- 36. glair.ai Machine Learning Workflow Business Problem ML Problem Framing Data Preparation EDA Feature Engineering Model Training and Tuning Model Evaluation Model Deployment Data Ingestion Are Business Goals Met? Data Augmentation Feature Augmentation No Yes 36
- 37. ● Model Serving/Inference ● Considerations: ○ How to wrap the prediction code as a production-ready service? ○ Which API / Protocol to use? ○ Scalability, Throughput, Latency. ○ Deployments ■ Model Versioning ■ Choose appropriate deployment strategy (e.g, A/B Testing) GDP LABS CONFIDENTIAL Model Deployment 07 37
- 38. glair.ai Machine Learning Workflow Business Problem ML Problem Framing Data Preparation EDA Feature Engineering Model Training and Tuning Model Evaluation Model Deployment Model Monitoring Data Ingestion Are Business Goals Met? Predictions Data Augmentation No Yes 38
- 39. ● Maintain model performance ● Things to be monitored: ○ Service Health ○ Data Quality & Integrity ○ Data & Target Drift ○ Bias/fairness GDP LABS CONFIDENTIAL Model Monitoring 08 39
- 40. Novan Parmonangan Simanjuntak Head of Machine Learning and AI Strategy novanps novan.p.simanjuntak@glair.ai glair glair.ai hi@glair.ai Contact Us Thank You!