Analytics machine learning in weka
Download as PPTX, PDF1 like748 views
WEKA is a collection of machine learning algorithms that can be used for data mining tasks. It supports various file formats like ARFF. The document discusses how to load data, perform univariate analysis, feature selection using info gain and correlation, and divide data into training, validation and test sets. It also demonstrates building a logistic regression model, analyzing it using ROC curves and cost-benefit analysis. WEKA allows reapplying models on new data and has advantages like being platform independent with a GUI, APIs and CLI, but has limitations such as limited visualizations and file formats.
Analytics machine learning in weka
- 2. Index • WEKA Introduction • WEKA file formats • Loading data • Univariate analysis • Data Manipulation • Feature Selection • Creating Training, Validation and Test Sets • Model Execution - Logistic Regression • Model Analysis - ROC Curve • Model Analysis – Cost/Benefit Analysis • Re-apply model on new data • Weka Plus and Negatives
- 3. Introduction • Weka is a collection of machine learning algorithms for data mining tasks • The algorithms can either be applied directly to a dataset or called from your own Java code. • ARFF – Attribute relation file format
- 7. Univariate Analysis • Current Relation – Dataset name, number of records, number of attributes in dataset. • Attribute details- All attributes to select for univariate analysis
- 8. Univariate Analysis • Selected attribute – • Provides information about attribute type, Missing values, Distinct values, etc.
- 9. Univariate Analysis • Selected attribute – histogram • Dispersion of attribute.
- 10. Univariate Analysis • All attribute visualization/plots
- 11. Data Manipulation • Changing data type of field • Missing values update • Creating BINS from data • Standardize data • Outlier Treatment • Creating new calculated fields
- 12. 1. Convert NA to 0 • Flow > Preprocess > Edit > Right Click on Attribute > Replace Values
- 13. 1. Convert NA to 0
- 14. 2. Changing data types of attribute • Flow > Preprocess > Filter > Choose > Filters > Unsupervised > Attribute >
- 15. 2. Changing data types
- 16. 3. Creating BINS • Flow > Preprocess > Filter > Choose > Filters > Unsupervised > Attribute > Discretize
- 17. 3. Creating BINS • Provide attribute number, Number of BINs to be created and click on ‘Apply’
- 18. 3. Creating BINS • Click on attribute to see the bins and distribution •
- 19. 3. Creating Custom BINS • Flow > Preprocess > Filter > Choose > Filters > Unsupervised > Attribute > AddExpression • ifelse(a2 > 0, ifelse(a2 > 10,ifelse(a2 > 20,4, 3), 2), 1)
- 20. 4. Standardize data • To convert all numeric attributes in data to zero mean and unit variance. • Flow > Preprocess > Filter > Choose > Filters > Unsupervised > Attribute > Standardize •
- 22. 4. Standardize data- Log values • To convert specific numeric attributes to log. • Flow > Preprocess > Filter > Choose > Filters > Unsupervised > Attribute > Numeric Transform •
- 23. 4. Standardize data- Log values • Provide value for attribute number which is to be converted to log value. • Also provide method name – log. Here we can provide any other methods such as abs,round,floor •
- 24. 5. Identify Outliers • Flow > Preprocess > Filter > Choose > Filters > Unsupervised > Attribute > Interquartile Range • Outliers can be identified for separate attribute or for all together
- 26. 5. Remove Outliers • Flow > Preprocess > Filter > Choose > Filters > Unsupervised > Instance > RemoveWithValues
- 27. 5. Remove Outliers • Params : attributeIndices - Attribute number, NominalIndices=Nominal value of outlier in Outlier attribute
- 28. 5. Transform Outliers • Flow > Preprocess > Filter > Choose > Filters > Unsupervised > Attribute > AddExpression – this option will create new field e.g : ifelse(a2 > 1000,200, 1) •
- 29. 6. New Calculated fields • This is helpful In case any new field is to be derived from existing fields • Flow > Preprocess > Filter > Choose > Filters > Unsupervised > Attribute > Add Expression
- 30. 6. New Calculated fields • This is helpful In case any new field is to be derived from existing fields • Provide expression/equation and new field name
- 31. Feature Selection/Attribute selection 1. Info Gain 2. Correlation
- 32. Feature Selection – Info Gain • Flow > Select Attribute > Attribute evaluator > Choose >
- 33. Feature Selection – Correlation • Flow > Select Attribute > Attribute evaluator > Choose >
- 34. Features for Model • Features selected for model
- 35. Creating Training, Validation and Test Sets
- 36. Creating data sets 1. Dividing data into 60-20-20 % (Train-Test-Evaluate) 2. Weka inbuilt methods
- 37. Creating data sets • For 60%-20%-20% • Step 1- • Flow > Preprocess > Filter > Choose > Unsupervised > Instance > Resample
- 38. Creating data sets • Step 2- Parameters for resample • Flow > Preprocess > Filter > Choose > Unsupervised > Instance > Resample > • Check noReplacement = True, sample size percent – 60 > ok > Apply
- 39. Creating data sets • Step 3 – Check • After apply we can check the current relation for number of records selected • • Step 4 – Save the result as filename_train.arff • Step 5 – Click on ‘Undo’ to get to original data set • Step 6 – Change the Resample parameters again • Parameters - > Invert selection = True, noReplacement = True, sampleSizePercent = 60
- 40. Creating data sets • Step 7 – Apply and check results as below
- 41. Creating data sets • Step 8 – Don’t save the results • Step 9 – Open Resample parameters set below parameters • Invert selection = False, noReplacement = True, sampleSizePercent = 50 • OK > Apply Check the results
- 42. Creating data sets • Step 10 – Check the results • Step 11 – Save the results as Test Data. • Step 13 – Click on Undo to get earlier 40% of dataset • Step 13 – Parameters invertSelection = True, noReplacement=True , sampleSizePercent = 50 • Step 14 – Ok > Apply • Step 15 – Save as Evaluation Data
- 43. Creating data sets • .
- 44. Creating data sets 2. Weka inbuilt – Flow > Classify > Test Options > Use Training Set – With this option selected data set will be used as training set to create model
- 45. Creating data sets 2. Weka inbuilt – Flow > Classify > Test Options > Use Supplied Test Set – With this option selected data set will be used as test set to create model
- 46. Creating data sets 2. Weka inbuilt – Flow > Classify > Test Options > Use Supplied Test Set – With this option selected data set will be used as test set to test model
- 47. Creating data sets 2. Weka inbuilt – Flow > Classify > Test Options > Cross Validation – With this option selected data set will be divided into 10 folds create model internally and weka will take average of all these models to show final model on UI
- 48. Creating data sets 2. Weka inbuilt – Flow > Classify > Test Options > Percentage Split – With this option selected data set will be divided into Training and Test set for model creation
- 50. Logistic Regression • Flow > Classify > Functions > Logistic
- 51. Logistic Regression • Parameter selection
- 52. Logistic Regression • Model Results:
- 53. Model Analysis • Flow > Right Click on Model > Visualize threshold curve > ROC Curve
- 54. Model Analysis- ROC Curve
- 55. Model Analysis- Cost Benefit Analysis • Flow > Classify > Right click model > Cost/Benefit Analysis
- 56. Model Analysis- Cost/Benefit Analysis • Flow > Classify > Right click model > Cost/Benefit Analysis > Threshold Bar • Sliding the bar under Threshold label will change the accuracy and threshold curve
- 57. Save prediction output to file • Flow > Classify > Test Options > More Options > Output Predictions > Text Bar to provide file name • Parameters: Choose: to provide file type, Attributes : First-last to get all fields, outPutFile : File name to save data
- 58. Re-apply model on new data
- 59. WEKA Pluses: • Platform independent and portable, java library can be invoked from any program in any language • User friendly GUI, with built in visualization, Simpler to use than R, large collection of different data mining algorithms • Better results for classification and cluster modeling • Ease of designing solutions. • Provides 3 ways to use the software: the GUI, a Java API, and a command line interface (CLI) • Can work with Spark, BigData using other packages on Experimenter or in batch mode.
- 60. WEKA Limitation: • Visualizations can be managed better in R with different packages like ggplot • Not really flexible for data manipulation • Accepts only limited file format lile CSV,ARFF • Limited documentation available on Explorer.
- 61. THANK YOU !
- 62. Decision Tree
- 63. Decision Tree: Algorithm selection
- 64. Decision Tree: Setting params for algo