Arm Treasure Data Internship 2018 Final Report

Transcript

1. Arm Treasure Data Summer Intern 2018 Final Report 2018/09/20 Masaki Kobayashi #td_intern
2. Who am I? 2 - Masaki Kobayashi (@_Makky_, @makky3939) - Master's student, University of Tsukuba - Research Topics: Crowdsourcing - Hobbies: JavaScript, Photography
3. About My Internship 3 - 2018/08/01 ~ 2018/09/28 - Mentors: @makimoto, @myui
4. What I did - I developed LDA-board - A topic model visualization tool using LDA - https://github.com/treasure-data/lda-board 4
5. Outline of This Talk Background & Related Projects What is LDA-board? Features & Architecture Evaluation & Conclusion 5
6. Outline of This Talk Background & Related Projects What is LDA-board? Features & Architecture Evaluation & Conclusion 6
7. Understanding Users from Collected Data • Can grasp what kind of people using our service • Can provide users oriented features e.g. Suggest/recommend something based on users data (or activity log) Background 7
8. LDA: Latent Dirichlet Allocation One of the most popular documents clustering method 8 Documents Topics A B C
9. LDA: Latent Dirichlet Allocation - LDA also can be applicable to any kind of clustering including user clustering. (not only documents) - TD users can be use LDA on Hive QL Hivemall has implementation of LDA - Already prepared a digdag workflow template to use LDA. 9
10. Treasure Data Customers Use Case 10 Users Topics A B C Sushi ! JavaScript Rails
11. Understanding each topic meaning - How were the documents divided? - Which topic contains this document (or user)? - What is the relationship between topics? (close? nested?) Topic Model Visualization 11
12. Related Projects 12 pyLDAvis [1] BigML [2] [1] http://nbviewer.jupyter.org/github/bmabey/pyLDAvis/blob/master/notebooks/pyLDAvis_overview.ipynb [2] https://bigml.com/shared/topicmodel/shSAsibY4TdQGIZvjjwtjs85cXx/topics can't check documents of each topics can't filter topics by term
13. There is no simple method to visualize hivemall's lda Of course, We can use pyLDAvis if we transform the output of hivemall Not all humans can use python / Jupyter notebook In Treasure Data users use cases, The tool is mainly used for user's analysis. They want to observe various information on users belonging to each topic. Issues 13
14. Development Requirements 1. Search and explore by term 2. Visualize clusters in two-dimensional space 3. Retrieve docids/userid in the specified topic 14
15. Outline of This Talk Background & Related Projects What is LDA-board? Features & Architecture Evaluation & Conclusion 15
16. LDA-board 16 - A topic model visualization tool using LDA - Focused on users data (e.g. activity log) - Can be easily visualize collected data in Treasure Data.
17. Quick start 0. Register your LDA workflow 1. Sign in with TD API Key 2. Run the workflow 3. Visualize 17
18. Quick start 0. Register your LDA workflow 1. Sign in with TD API Key 2. Run the workflow 3. Visualize 18
19. Quick start 0. Register your LDA workflow 1. Sign in with TD API Key 2. Run the workflow 3. Visualize 19
20. Quick start 0. Register your LDA workflow 1. Sign in with TD API Key 2. Run the workflow 3. Visualize 20
21. Quick start 0. Register your LDA workflow 1. Sign in with TD API Key 2. Run the workflow 3. Visualize 21
22. Outline of This Talk Background & Related Projects What is LDA-board? Features & Architecture Evaluation & Conclusion 22
23. Features of LDA-board 1. Manage workflow executions 2. Run a workflow with setting params 3. Fetch the result of workflow from TD 4. Visualize the result 23
24. Manage workflow executions 24
25. Run a workflow with setting params 25
26. based on https://github.com/treasure-data/workflow-examples/tree/master/machine-learning/lda Task (lda.dig) + tokenize + prepare_input_table + train_lda + post_train Run a workflow with setting params 26
27. https://github.com/treasure-data/workflow-examples/tree/master/machine-learning/lda Task (lda.dig) _export: !include : config/params.yml num_topics: "${typeof(session_num_topics) === 'undefined' ? default_num_topics : session_num_topics}" + tokenize + prepare_input_table + train_lda + post_train Run a workflow with setting params 27
28. https://github.com/treasure-data/workflow-examples/tree/master/machine-learning/lda Task (lda.dig) + tokenize + prepare_input_table + train_lda + post_train + post_predict Run a workflow with setting params 28 + dimension_reduction + topic_proportion + prepare_pca_input: + run_pca: py>: tasks.DimensionReduction.jspca docker: image: "python:3.6.5" _env: python_apikey: ${secret:python_apikey}
29. Fetch the result of workflow from TD 29
30. Visualize the result 30
31. Visualize the result - Filter Topics by term 31
32. Show topic terms and elements 32
33. Additional contents in Topic elements 33
34. Additional contents in Topic elements 34
35. Architecture 35 Amazon ECS API Server (Rails) Browser Client App (React / Redux) Amazon RDS PostgreSQL https://api.treasuredata.com https://api-workflow.treasuredata.com
36. Try with local docker-compose 36
37. Outline of This Talk Background & Related Projects What is LDA-board? Features & Architecture Evaluation & Conclusion 37
38. Evaluation - Development Requirements - Feedback from TD Customers (LDA users) 38
39. Development Requirements 1. Search and explore by term 2. Visualize clusters in two-dimensional space 3. Retrieve docids/userid in the specified topic 39
40. Feedback from TD Customers (LDA users) - Want to use <Link> in Additional contents area e.g. To users detail page - Want to show Additional contents in Topic terms e.g terms -> stores - Just run post_predict part to avoid lda train time 40
41. Future Tasks (want to try Final Week) 41 - Prepare for publication - Documentation - Add more examples for expected use cases
42. Conclusion 42 - I developed LDA-board - A topic model visualization tool using LDA - github.com/treasure-data/lda-board
43. Acknowledgements 43 Special thanks to … @makimoto -san @myui -san and All Treasure Data members