Is it Corked? Wine Machine Learning Predictions with OAC

Transcript

2. info@rittmanmead.com www.rittmanmead.com @rittmanmead Is It Corked? Wine Machine Learning Predictions with OAC Francesco Tisiot BI Tech Lead at Rittman Mead
3. Verona, Italy http://ritt.md/ftisiot 10 Years in Analytics francesco.tisiot@rittmanmead.com @FTisiot Oracle ACE Francesco Tisiot BI Tech Lead at Rittman Mead
4. Specialised in data visualisation, predictive analytics, enterprise reporting and data engineering. Enabling the business, the consumers, the data providers and IT to work towards a common goal, delivering innovative and cost-effective solutions based on our core values of thought leadership, hard work and honesty. We work across multiple verticals on projects that range from mature, large scale implementations to proofs of concept and can provide skills in development, architecture, delivery, training and support. www.rittmanmead.com info@rittmanmead.com @rittmanmead
5. Do You Like Alfredo’s Fettuccine ?
6. Who is Alfredo ?
7. info@rittmanmead.com www.rittmanmead.com @rittmanmead Agenda !7 •Tooling •Data Science Steps •Demo
8. info@rittmanmead.com www.rittmanmead.com @rittmanmead Tooling
9. info@rittmanmead.com www.rittmanmead.com @rittmanmead !9 Oracle Analytics Cloud • Oracle’s complete suite of Platform Services (PaaS) for unified analytics in the cloud • Delivered entirely in the cloud: ‣ No infrastructure footprint ‣ Flexibility to scale up or down based on your immediate needs ‣ Simplified, metered licensing • Several options to suit your needs: ‣ Oracle or customer/partner managed services ‣ Functionality bundled into 3 editions
10. info@rittmanmead.com www.rittmanmead.com @rittmanmead • OAC supports every type of analytics workload across your organisation !10 Functions • Classic enterprise BI: ‣ Analysis & dashboarding ‣ Published reporting ‣ Enterprise Performance Management • Modern departmental/personal discovery: ‣ Extended data mashup & modelling ‣ Data preparation, exploration & visualisation ‣ Data science & machine learning
11. info@rittmanmead.com www.rittmanmead.com @rittmanmead • Similar User Experience to OBIEE 12c ‣ Centrally maintained & governed ‣ Semantic model remains key • Interactive Dashboards ‣ Ideal for KPI measurement & monitoring ‣ Guided navigation paths • BI Publisher ‣ Highly formatted, burst outputs • Action Framework ‣ Navigation actions ‣ Scheduled agents !11 Classic Enterprise BI
12. info@rittmanmead.com www.rittmanmead.com @rittmanmead • Data Preparation ‣ Acquire data from multiple connections ‣ Apply enrichments data prior to analysis ‣ Define repeatable preparation flows • Data Visualisation ‣ Create visual insights rapidly ‣ Construct narated storyboards ‣ Share findings • Machine Learning ‣ Build & train ML models ‣ Apply model to new data sets !12 Modern Data Discovery
13. info@rittmanmead.com www.rittmanmead.com @rittmanmead OAC And Data Science
14. Basic Operations What are the Drivers for My Sales? Based on my Experience I can Guess…. Statistically Significant Drivers for Sales Are … Augmented Analytics
15. Basic Operations Is this Client going to accept the Offer? YES/NO 50% 70% Basic ML Model
16. Before Starting…. Define the Problem!
17. Problem Definition: Predicting Wine Quality
18. Rule Based Italy or France -> Good Rest of the World -> Bad Price >= 10 Euros -> Good Price < 10 Euros -> Bad Price > 30 & Production Zone = Veneto & …. -> 6.5
19. Task Experience Performance Estimate Wine Good/Bad TEP Corpus of Wines Descriptions with Ratings Accuracy
20. Accuracy Icons made by Smashicons from www.flaticon.com Real Value Predicted Value Good Bad Bad Good / ( ) + Accuracy =
21. Dataset https://www.kaggle.com/zynicide/wine-reviews
22. The Data
23. Bad Good
24. Become a Data Scientist with OAC Connect Clean Analyse Train & Evaluate Predict Transform & Enrich
25. Connect Pre-Defined Data Models Data Sources
26. Clean N/A Missing Values Mark <> MArk Wrong Values City “Rome” Irrelevant Observations Role: CIO Salary:500 K$ Handling Outliers Train: 80% Test: 20% Train/Test Set Split Col1 -> Name Labelling Columns 0-200k 0-1 Feature Scaling # Of Clicks Aggregation
27. Why Removing an Outlier? Years Experience Salary 1 30.000 2 32.000 3 35.000 4 35.500 5 36.000 6 40.000 7 50.000 8 70.000 9 90.000 10 500.000
28. Enrich - Feature Engineering Location -> ZIP Code 2 Locations -> Distance Name -> Sex Day/Month/Year -> Date Data Flow Additional Data Sources?
29. Data Preparation Recommendations
30. Analyse - Data Overview
31. Analyse - Explain
32. Explain - Key Drivers
33. Train - What Problem are we Trying to Solve? Supervised Unsupervised “I want to predict the value of Y, here are some examples” “Here is a dataset, make sense out of it!” Classification Regression https://towardsdatascience.com/supervised-vs-unsupervised-learning-14f68e32ea8d Clustering
34. Easy Models
35. DataFlow Train Model
36. Which Model - Parameters To Pick?
37. Select, Try, Save, Change, Try, Save …..
38. Compare
39. Compare
40. There is No Single Truth… 502/(502+896) = 64.09% 471/(471+866)=64.77%
41. Predict - Use On the Fly
42. Predict - Step of a Data Flow
43. Demo
44. Conclusions 73% > 63% > 50% Existing Skillset Trial Error Visual -> UI Driven Data Cleaning & Transformation Model Creation & Evaluation
45. Example https://towardsdatascience.com/wine-ratings-prediction-using-machine-learning-ce259832b321
46. Custom ML Models Model_train.xml Parameter Definition Python Parameter Parsing Data Cleaning Model Storage Statistics Calculation svr=SVR(kernel=kernel, gamma=0.01, C= 5) SVR_Model = svr.fit(train_X, train_y) Model Definition & Training Model_test.xml https://www.oracle.com/solutions/business-analytics/data-visualization/library-overview.html
47. Custom ML - Parameter Definition <options> <option> <name>target</name> <displayName>Target</displayName> <value></value> <required>true</required> <description>target, the target(label) to learn/predict</description> <type>column</type> <ui-config></ui-config> </option> <option> <name>description</name> <displayName>Description</displayName> <value></value> <required>true</required> <description>The descriptive field to be parsed</description> <type>column</type> <ui-config></ui-config> </option> </options>
48. Custom ML - Script <scriptcontent> </scriptcontent> Import Libraries def obi_create_model(…)
49. Parameter Parsing target = args['target'] description = args['description'] test_size = float(args['split'])
50. train_X, test_X, train_y, test_y = train_test_split( X, y, test_size=test_size, random_state=0 ) Train-Test Split
51. Model Creation rfc = RandomForestClassifier() rfc.fit(train_X, train_y) predictions = rfc.predict(test_X)
52. Model Storage pickleobj={'WineModel':rfc} d = base64.b64encode(pickle.dumps(pickleobj)).decode(‘utf-8') required_mappings = {'description':'true'} model = Model() model.set_data(__name__, 'pickle', d, target) model.set_required_attr(required_mappings) model.set_class_tag('BinaryClassification') target_col=y
53. Calculating Confusion Matrix # predict test_data with model predictions_proba = rfc.predict_proba(test_X) predictions = lrutils.get_classicification(predictions_proba, rfc.classes_, threshod,positiveValue, targetMappings) # confusion matrix dataset and statistics confusion_df, statistics_df = datasetutils.prepare_output_datasets(test_y, predictions, target, "classification", targetMappings) confusion_mappings = None confusion_ds = ModelDataset("Confusion Matrix", confusion_df, confusion_mappings) model.add_output_dataset(confusion_ds) # statistics dataset statistics_mappings = None statistics_ds = ModelDataset("Statistics", statistics_df, statistics_mappings) model.add_output_dataset(statistics_ds)
54. Result
55. Limitations • Libraries • Testing • Performances
57. info@rittmanmead.com www.rittmanmead.com @rittmanmead Is It Corked? Wine Machine Learning Predictions with OAC Francesco Tisiot BI Tech Lead at Rittman Mead