AIRLINE_SATISFACTION_Data Science Solution on Azure

Airline Satisfaction Project using Azure

This presentation is created as a foundation of understanding and comparing data science/machine learning solutions made in Python notebooks locally and on Azure cloud, as a part of Course DP-100 - Designing and Implementing a Data Science Solution on Azure.

1. July 2024 Sanela Nikodinoska AIRLINE SATISFACTION DATA SCIENCE SOLUTION ON

2. Agenda Introduction Automated ML Designer Notebooks – Python SDK Closing

3. Introduction For the last but most significant course DP – 100 – Designing and Implementing a Data Science Solution on Azure, part of Data Science Institute held by Semos Education, Airline Satisfaction dataset was given to design and implement a data science solution on Azure. This presentation is an overview of the implemented solutions created using Azure Machine Learning Studio. Since the subscription to Azure was made for learning purposes only and is now cancelled, this presentation is made upon screenshots of the most important steps while developing, training and deploying ml models.

4. Automated ML Screenshots from Azure Let’s dive in

5. First steps Started with an Azure free trial, created resource group from UI and created Azure Machine Learning Service

6. Data – created data asset, uploading local dataset Airline Satisfaction to Azure (no screenshot for that) Automated ML – created two experiments, setting different primary metrics and featurization parameters

7. Automated ML – best model in both experiments was MaxMinScaler , LightGBM , experiments stopped due to early stopping policy based on level of primary metric

8. Automated ML – metrics

9. Designer Screenshots from Azure Let’s dive in

10. Ju Authoring Using components from Authoring - > Designer tab, CREATED two pipelines with two estimators and one pipeline with single estimator model for feature importance component: Two-Class Logistic Regression and Two-Class Decision Forest, Two-Class Support Vector Machine and Two-Class Neural Network, (deep-learning model), Two-Class Boosted Decision Tree with Cross – Validate Model component Jobs / Metrics After configuring and submitting pipelines and images for env, a job was created. The overview of job, as well as its outputs, logs, child jobs and metrics are presented in the following snapshots Registered Models The best performing models from Automated ML and Designer (Neural Network and LightGBM ) registered as custom and mflow models Real-time Endpoint Blue/green deployment of two best models was made and the blue deployment was tested for inference / the endpoint was invoked DESIGNER Compute targets Compute instance for profiling data asset, compute clusters for training models and pipeline sweep jobs were created Environments Compute instance for profiling data asset, compute clusters for training models and pipeline sweep jobs were created

11. DESIGNER Two-Class Logistic Regression and Two-Class Decision Forest pipeline

12. DESIGNER Two-Class Support Vector Machine and Two-Class Neural Network

13. DESIGNER Two-Class Decision Tree with Feature Importance component and Cross – Validate Model component

14. Designer - Feature importance

15. JOBS – list of all experiments

16. JOBS – overview of designer pipelines metrics – Random Forest Classifier best model Note: No snapshots of pipelines success

17. JOBS – overview of designer pipelines metrics –Two-Class Neural Network model Note: No snapshots of pipelines success

18. JOBS – overview of designer pipelines metrics –Two-Class Logistic Regression least performing model Note: No snapshots of pipelines success

19. Registered models

20. Deployment of models

21. Real-time Endpoint

22. Custom environment for model deployed to the endpoint

23. Predicting / Invoking the Endpoint

24. Environments Custom and curated environments for deploying and testing

25. Compute targets

26. Notebooks Screenshots from Azure Let’s dive in

27. Notebooks – created pipeline for training and scoring RandomForestClassifer model and tunning hyperparameters with sweep job

28. Notebooks – running tunning hyperparameters with sweep job

29. Notebooks – results from pipeline – child jobs

30. Notebooks – results from pipeline – best model

31. Notebooks – results from pipeline – model metrics

32. Notebooks – results from pipeline - predicting

33. Summary No-code or programmatically What suits the most Great business solution Having all resources in one place New subscription For future projects Getting work done F inished my course project So many services Yet to discover: Azure DataBricks , Azure Synapce Analystics (for data ingestion), Azure AI Services, Azure Data factory etc. Recommend Definitely!

34. Closing Thanks to your time. Hoping to get some of your feedback for improving. Sanela Nikodinoska [email protected]