MLSD18. Unsupervised Workshop
1 like135 views
This document summarizes exercises from a machine learning school workshop on unsupervised models using the BigML platform. It includes instructions for attendees to complete exercises on anomaly detection, clustering, topic modeling and association discovery on several healthcare and other datasets. The exercises are meant to demonstrate how to build and interpret different unsupervised learning models in BigML.
MLSD18. Unsupervised Workshop
- 1. 1st edition November 4-5, 2018 Machine Learning School in Doha
- 2. BigML, Inc · @bigmlcom · @QatarComputing · #MLSD18 · Build Your Own Unsupervised Models Gregory Antell, Ph.D. Machine Learning Architect, BigML Inc. !2 The impact of hyperparameter choices
- 3. BigML, Inc · @bigmlcom · @QatarComputing · #MLSD18 · https://bigml.com/gallery/datasets !3 Professors Salaries: http://bml.io/13s39KL (search under “Higher Education and Research”) Reading Habits: http://bml.io/XGXCg3 (search under “Demographics and Surveys”) TED Talks: https://bigml.com/user/czuriaga/gallery/dataset/ 5ab8fabc92fb562acb00170a (search under “Human Resources and Psychology”) Diabetes: http://bml.io/QdJPdb (search under “Healthcare”)
- 4. BigML, Inc · @bigmlcom · @QatarComputing · #MLSD18 · Unsupervised Learning !4 • Unsupervised methods seek to identify and uncover the hidden structure of complicated datasets • In general, these models require higher levels of domain knowledge and interpretation to in order to validate • Selection of features and initialization can have a massive impact on the results of unsupervised algorithms
- 5. BigML, Inc · @bigmlcom · @QatarComputing · #MLSD18 · BigML Unsupervised Options !5 ANOMALY DETECTION ASSOCIATION DISCOVERY CLUSTER TOPIC MODEL
- 6. BigML, Inc · @bigmlcom · @QatarComputing · #MLSD18 · Exercise 1 - Anomaly Detection !6 Professors salaries dataset: http://bml.io/13s39KL (search under “Higher Education and Research”) • Clone the dataset • Create an anomaly detector returning at least the top 10 anomalous instances • What is the gender, salary, and years since Ph.D for the top anomaly? • What is the rank and years of service for the #2 anomaly
- 7. BigML, Inc · @bigmlcom · @QatarComputing · #MLSD18 · Exercise 1 - Anomaly Detection !7 Professors salaries dataset: http://bml.io/13s39KL (search under “Higher Education and Research”) • Clone the dataset • Create an anomaly detector returning at least the top 10 anomalous instances • What is the gender, salary, and years since Ph.D for the top anomaly? • What is the rank and years of service for the #2 anomaly Female, $62884, 25 Associate Professor, 53
- 8. BigML, Inc · @bigmlcom · @QatarComputing · #MLSD18 · Exercise 2 - Clustering !8 • Clone the dataset • Create a cluster using the G-means option • How many different clusters are returned? • Repeat but remove race, marital status, and fields regarding newspapers and magazines. How many clusters now? Reading Habits: http://bml.io/XGXCg3 (search under “Demographics and Surveys”)
- 9. BigML, Inc · @bigmlcom · @QatarComputing · #MLSD18 · Exercise 2 - Clustering !9 • Clone the dataset • Create a cluster using the G-means option • How many different clusters are returned? • Repeat but remove race, marital status, and fields regarding newspapers and magazines. How many clusters now? Reading Habits: http://bml.io/XGXCg3 (search under “Demographics and Surveys”) 8 2
- 10. BigML, Inc · @bigmlcom · @QatarComputing · #MLSD18 · Exercise 3 - Topic modeling !10 • Clone the dataset • Create a topic model using the default options and only the fields “title” and “description • How many different topics are returned? • What are some other terms highly associated with topics that deal with healthcare? TED Talks: https://bigml.com/user/czuriaga/gallery/dataset/ 5ab8fabc92fb562acb00170a (search under “Human Resources and Psychology”)
- 11. BigML, Inc · @bigmlcom · @QatarComputing · #MLSD18 · Exercise 3 - Topic modeling !11 • Clone the dataset • Create a topic model using the default options and only the fields “title” and “description • How many different topics are returned? • What are some other terms highly associated with topics that deal with healthcare? TED Talks: https://bigml.com/user/czuriaga/gallery/dataset/ 5ab8fabc92fb562acb00170a (search under “Human Resources and Psychology”) 40 topics Research, disease, cancer, drug, etc.
- 12. BigML, Inc · @bigmlcom · @QatarComputing · #MLSD18 · !12 • Clone the dataset • Create an association discovery • Sort associations by lift. What features are compared in the top association? • Sort associations by leverage. Do the top feature associations change? Diabetes: http://bml.io/QdJPdb (search under “Healthcare”) Exercise 4 - Associations
- 13. BigML, Inc · @bigmlcom · @QatarComputing · #MLSD18 · !13 • Clone the dataset • Create an association discovery • Sort associations by lift. What features are compared in the top association? • Sort associations by leverage. Do the top feature associations change? Diabetes: http://bml.io/QdJPdb (search under “Healthcare”) Exercise 4 - Associations Yes: triceps skin thickness and insulin triceps skin thickness and BMI