Data Mining 101

Introduction to data mining based on CRISP DM

Transcript

1. Data Mining 101 Okiriza Wibisono - @okiriza Ali Akbar Septiandri - @aliakbars
2. Outline Introduction •Terminology •Potential application •Venn diagram Process overview •Business understanding •Data understanding (exploration) •Data preparation (preprocessing) •Modeling •Evaluation •Deployment (presentation) Tools & Resource
3. Introduction – Terminology Data mining Knowledge Discovery in Databases Big data analytics Statistics Data science
4. “ ” The process of collecting, searching through, and analyzing a large amount of data in a database, as to discover patterns or relationships. Data Mining - dictionary.reference.com
5. Introduction – Potential Application Customer segmentation Recommendation engine Social media mining
6. “ ” What should we do? Where to start? Do I have to get a master degree in statistics?
7. http://tomfishburne.com.s3.amazonaws.com/site/wp-content/uploads/2014/01/140113.bigdata.jpg
8. Data Science Venn Diagram http://drewconway.com/zia/2013/3/26/the- data-science-venn-diagram
9. And now the business process…
10. CRISP DM Methodology http://lyle.smu.edu/~mhd/8331f03/crisp.pdf
11. Business Understanding CRISP DM Methodology
12. Objective Statement Bottom-up Top-down
13. Objective Statement Data Problem vs
14. Situation Assessment Inventory of Resources Requirements, Assumptions, and Constraints Risks and Contingencies Terminology Costs and Benefits http://www.ismll.uni-hildesheim.de/lehre/ba-12ss/script/ba03.pdf
15. Situation Assessment – Inventory of Resources Resource Data, Knowledge, Tools Hardware Personnel http://www.ismll.uni-hildesheim.de/lehre/ba-12ss/script/ba03.pdf
16. Situation Assessment – Requirements, Assumptions, and Constraints Requirements Scheduling Accuracy Security Assumptions Data quality External factors Reporting type Constraints Legal issues Budget Resources http://www.ismll.uni-hildesheim.de/lehre/ba-12ss/script/ba03.pdf
17. Situation Assessment – Risks and Contingencies Contingency Plan Financial Organizational Business http://www.ismll.uni-hildesheim.de/lehre/ba-12ss/script/ba03.pdf
18. Situation Assessment – Terminology Write down related terminology http://www.ismll.uni-hildesheim.de/lehre/ba-12ss/script/ba03.pdf http://www.partnersmn.com/wp-content/uploads/2010/08/5b8567b2b4e2d1cfd1a31b2b8a0ecebc1.jpg
19. Situation Assessment – Costs and Benefits Money, money, money! http://www.ismll.uni-hildesheim.de/lehre/ba-12ss/script/ba03.pdf http://www.centuryproductsllc.com/wp-content/uploads/holding-money.jpg
20. “ ” How to evaluate the results? Define your success criteria!
21. Data Understanding CRISP DM Methodology
22. Data Collection External Internal vs
23. Watch out!
24. “ ” visible ≠ accessible ≠ storable ≠ presentable Victor Lavrenko – Text Technologies http://www.inf.ed.ac.uk/teaching/courses/tts/pdf/crawl-2x2.pdf
25. Data Exploration – Visualization Heuristics  Visualize fast. Visualize reactively.  Go for high information 2D visualizations.  Select data subsets to visualize. http://www.inf.ed.ac.uk/teaching/courses/dme/slides2014/visualization-print4up.pdf
26. Data Exploration – Visualization Heuristics  Never let anomalies pass you by. Dig deeper.  Use your visualizations to inform potential models. Use your potential model to direct your visualizations.  Expect problems in your data. http://www.inf.ed.ac.uk/teaching/courses/dme/slides2014/visualization-print4up.pdf
27. “ ” This is the cheapest and most informative stage of data mining. Nigel Goddard – DME Visualization http://www.inf.ed.ac.uk/teaching/courses/dme/slides2014/visualization-print4up.pdf
28. Data Exploration – Visualization Tools  Column/bar: Large change  Line, curve: Small change, long periods  Histogram: Frequency distribution https://nces.ed.gov/nceskids/help/user_guide/graph/whentouse.asp
29. Data Preparation CRISP DM Methodology
30. “ ” Which one should I include (or exclude)? Data Selection
31. Data Cleaning Dirty Data Missing value Incomplete Outdated Duplication Outlier Remember: Expect problems in your data.
32. Data Construction  Feature engineering – derived attributes, e.g.: year from timestamp quarter from timestamp BMI from weight and height Log(x) for skewed data (e.g. house price)
33. Data Splitting Two kinds of data splitting: Training-Validation-Testing Cross Validation
34. Data Splitting – Training-Validation-Testing • Construct classifier Training • Pick algorithm • Knob settings (tree depth, k in kNN, c in SVM) Validation • Estimate future error rate Testing Split randomly to avoid bias http://www.inf.ed.ac.uk/teaching/courses/iaml/slides/eval-2x2.pdf
35. Data Splitting – Cross Validation Every point is both training and testing, never at the same time
36. Dimensionality Reduction Principal Component Analysis Linear Discriminant Analysis vs
37. Modeling CRISP DM Methodology
38. Machine Learning Classification Regression Ranking Clustering
39. Model Selection Regression Technique Generalization bound Linear regression Kernel ridge regression Support vector regression Lasso
40. “ ” Which one should I choose? Should I use all of them?
41. It depends on…
42. Model Selection Assumptions The predictors are linearly independent The error is a random variable with a mean of zero conditional on the explanatory variables The sample is representative of the population for the inference prediction Interpretability The understandability of why the model is true or how the model is induced from https://chenhaot.com/pubs/mldg-interpretability.pdf
43. Beware of Overfitting! http://pingax.com/wp-content/uploads/2014/05/underfitting-overfitting.png
44. Model Assessment Regression • (R)MSE • Mean Absolute Error • Correlation Coefficient Classification • Accuracy • Precision • Recall • F-score Descriptive • Std. Error • p-value • Confidence Interval
45. Evaluation CRISP DM Methodology
46. “ ” Does my model solve the problem? What is the impact? Is it novel? How useful is the solution?
47. Deployment CRISP DM Methodology
48. The Tasks Plan deployment Plan monitoring and maintenance Produce final report Review project
49. Tools & Resource  Text mining: NLTK, spaCy, OpenNLP  Query expansion & clustering: Carrot2, Weka  Data mining & machine learning: Weka, scikit-learn  Language: R, Python, Julia, Java, Matlab, Mathematica, Haskell, Scala  Python lib: Pandas, SciPy, NumPy, scikit-learn  Infrastructure: AWS, Hadoop, Google Cloud, Azure, Apache Spark  Visualization: D3.js  Community: Big Data & Open Data Indonesia
50. “ ” Thank you! Data Mining 101 – Python-ID Meetup February 2015 Okiriza Wibisono - @okiriza Ali Akbar Septiandri - @aliakbars