Cluster, Classify, Associate, Regress: Satisfy Your Inner Data Scientist with OML and Analytics
Download as PPSX, PDF0 likes60 views
The document discusses machine learning (ML) techniques and their applications in the context of solar energy investment and customer analytics. It outlines methods such as clustering, classification, association, and regression to analyze customer data and predict behaviors, aiming to leverage insights for better targeting and service offerings. Additionally, it highlights advancements in automated ML (AutoML) tools that simplify the algorithm selection process for users without extensive data science backgrounds.
Cluster, Classify, Associate, Regress: Satisfy Your Inner Data Scientist with OML and Analytics
- 1. Cluster, Classify, Associate, Regress: Satisfy Your Inner Data Scientist with OML and Analytics Jim Czuprynski @JimTheWhyGuy Zero Defect Computing, Inc. Photo credit: National Cancer Institute on UnSplash
- 2. Who Am I, and How Did I Get Here? E-mail me at jczuprynski@zerodefectcomputing.com Follow me on Twitter (@JimTheWhyGuy) Connect with me on LinkedIn (Jim Czuprynski) Traveler & public speaker Summers: Wisconsin Winters: Illinois Cyclist XC skier Avid amateur bird watcher Oldest dude in Krav Maga class
- 3. Jim Czuprynski Liron Amitzi https://www.beyondtechskills.com The podcast that talks about everything tech – except tech.TM
- 4. Typical Business Use Cases Driving Machine Learning & Analytics Are any of our customers at financial risk to default on their investments in solar panel technology? Are current solar energy program customers likely to recommend our program to colleagues, or maybe even request more alternative energy sources? Has the 2020 pandemic hurt our best customers’ ability to continue to invest in alternate energy sources? What’s the long-term return our customers can expect to realize on their solar energy investment? Which customers would be best candidates to take advantage of our solar panel program?
- 5. ML Basics: Clustering, Classification, Association, Regression Which customers would be best candidates to take advantage of our solar panel program? Are any of our customers at financial risk to default on their investments in solar panel technology? What’s the long-term return our customers can expect to realize on their solar energy investment? Are current solar energy program customers likely to recommend our program to colleagues, or maybe even request more alternative energy sources? Clustering identifies how certain like items are grouped together into one or more previously undefined categories Classification lets us predict the probability for a single item to belong to a defined category, or even multiple categories Regression gives us the power to project trends forward in time within upper and lower confidence boundaries Association discovers the chance that two or more items are likely to be found within the same collection as well as the relationships between those items Check out this summary that breaks out all of the inherent algorithms that support these methods
- 6. SIMIOT: A Modern Electric Utility’s Information Model SMART_METERS Contains information about individual smart meters for which data is being collected METER_READINGS Individual readings for each smart meter over (often extremely!) short time intervals DISPATCH_CENTERS Utility repairs depots from which servicepeople & trucks are dispatched BUSINESS_DESCRIPTIONS Describes unique business classifications based on licensing issued CUSTOMER_DEMOGRAPHICS Demographics for each customer, including ethnicity, minority status, and employee counts CUSTOMER_FINANCIALS Pertinent financial data for each customer, including credit worthiness CUSTOMER_RESPONSES Selected social media posts reflecting how customers anonymously view services & products
- 7. Probing Data Sources, Wherever They Live And What They Comprise Captured from an OSS stream - 10K readings every minute Most recent data within database, but history available as CSV files in Amazon S3 bucket (which cannot be moved!) Stored in an Excel spreadsheet … “somewhere” Pulled from City of Chicago public records via API calls, saved as CSVs and (re)loaded quarterly Captured via Twitter API stream, stored as multiple JSON documents
- 8. Converged Database: A Vision for the Future, 21c and Beyond Personal / External Datasets Enterprise Applications Data Integration OAC, OML, APEX, and Graph Studio Ad hoc, Batch or Scheduled Business Leaders Analysts Data Scientists Developers OAC Dataflow, Manual or ETL Data Management ADW Business Analytics ERP CRM HCM Self-sufficient, encrypted, secured data storehouse Self-service analytics via ML REST-Enabled External APIs IoT and Edge Computing ATP AJD AGD The new kid on the block: Autonomous JSON Database Recently announced: Autonomous Database for Graph Studio AutoML features make it simple to apply the right algorithm(s) with confidence
- 9. Picking the Right Algorithm: An Algorithm Would Be Nice … Here’s the good news: There’s at least one powerful Machine Learning algorithm to choose from for any analytic scenario Here’s the bad news: You’ve got to pick the right one considering its advantages and drawbacks for each scenario
- 10. Uncover Associations From Customer Survey Results (1) CS_ID CS_POSITIVE_RESPONSE ---------- ------------------------- . . . 2633595 PlanningToRenew 2633595 WindTurbineInterest 2635273 PlanningToRenew 2635273 RecommendedToColleague 2635273 WindTurbineInterest 2635712 PlanningToRenew 2637099 PlanningToRenew 2637464 PlanningToRenew 2637464 WindTurbineInterest 2637995 PlanningToRenew 2637995 RecommendedToColleague 2637995 WindTurbineInterest . . . Our marketing department recently surveyed the top 100 users of our solar power initiative. Some were excited enough to recommend it to others, and some even want to expand to wind turbines … … and others, not so much.
- 11. Uncover Associations From Customer Survey Results (2) DECLARE -- Processing variables: SQLERRNUM INTEGER := 0; SQLERRMSG VARCHAR2(255); vcReturnValue VARCHAR2(256) := 'Failure!'; v_setlist DBMS_DATA_MINING.SETTING_LIST; BEGIN v_setlst('ALGO_NAME’) := 'ALGO_APRIORI_ASSOCIATION_RULES'; v_setlst('PREP_AUTO’) := 'ON'; v_setlst('ASSO_MIN_SUPPORT') := '0.04'; v_setlst('ASSO_MIN_CONFIDENCE') := '0.1'; v_setlst('ASSO_MAX_RULE_LENGTH'):= '2'; v_setlst('ODMS_ITEM_ID_COLUMN_NAME'):= 'CS_POSITIVE_RESPONSE'; . . . We’ll use the APriori algorithm to uncover any hidden relationships between different responses to our recent customer satisfaction survey …
- 12. Uncover Associations From Customer Survey Results (3) . . . DBMS_DATA_MINING.CREATE_MODEL2( MODEL_NAME => 'CUSTOMER_RENEWABILITY' ,MINING_FUNCTION => 'ASSOCIATION' ,DATA_QUERY => 'SELECT * FROM t_survey_results' ,SET_LIST => v_setlst ,CASE_ID_COLUMN_NAME => 'CS_ID'); EXCEPTION WHEN OTHERS THEN SQLERRNUM := SQLCODE; SQLERRMSG := SQLERRM; vcReturnValue := 'Error creating new OML4SQL model: ' || SQLERRNUM || ' - ' || SQLERRMSG; DBMS_APPLICATION_INFO.SET_MODULE(NULL,NULL); DBMS_OUTPUT.PUT_LINE(vcReturnValue); END; / … using each customer’s ID as the individual case ID for the model
- 13. Uncover Associations From Customer Survey Results (4) SELECT * FROM ( SELECT rule_id ,antecedent_predicate antecedent ,consequent_predicate consequent ,ROUND(rule_support,3) supp_rtg ,ROUND(rule_confidence,3) conf_rating ,NUMBER_OF_ITEMS num_items FROM dm$vacustomer_renewability ORDER BY rule_confidence DESC, rule_support DESC ) WHERE ROWNUM <= 10 ORDER BY rule_id, antecedent, consequent; RULE_ID ANTECEDENT CONSEQUENT SUPP_RTG CONF_RATING NUM_ITEMS ------- ---------------------- ------------------------ -------- ----------- ---------- 1 RecommendedToColleague PlanningToRenew .172 .895 2 2 PlanningToRenew RecommendedToColleague .172 .179 2 3 WindTurbineInterest PlanningToRenew .818 .953 2 4 PlanningToRenew WindTurbineInterest .818 .853 2 5 WindTurbineInterest RecommendedToColleague .162 .188 2 6 RecommendedToColleague WindTurbineInterest .162 .842 2 A very simple query against the results of the Association model shows the individual support and confidence ratings for each survey response
- 14. Clustering, Brute Force: Looking for Solar Power Super-Users (1) CREATE OR REPLACE VIEW solar_superusers AS SELECT sm_business_type ,sm_zipcode ,sm_id ,ROUND(AVG(smr_kwh_used),2) avg_kwh_used ,ROUND(AVG(smr_solar_kwh),2) avg_solar_kwh ,ROUND(AVG(smr_solar_kwh) / AVG(smr_kwh_used) ,2) pct_solar ,CASE WHEN ROUND(AVG(smr_solar_kwh) / AVG(smr_kwh_used) ,2) >= 0.15 THEN 1 ELSE 0 END as solar_superuser FROM t_smartmeters ,t_meter_readings WHERE smr_id = sm_id GROUP BY sm_business_type, sm_zipcode, sm_id ORDER BY sm_business_type, sm_zipcode, sm_id ; We’ll use this metric to identify which SmartMeter customers are solar energy super-users (defined as 15% or more of their energy generated via solar power on-site)
- 15. Clustering, Brute Force: Looking for Solar Power Super-Users (2) DECLARE -- Processing variables: SQLERRNUM INTEGER := 0; SQLERRMSG VARCHAR2(255); vcReturnValue VARCHAR2(256) := 'Failure!'; v_setlist DBMS_DATA_MINING.SETTING_LIST; BEGIN v_setlist('ALGO_NAME') := 'ALGO_KMEANS'; v_setlist('PREP_AUTO') := 'ON'; v_setlist('KMNS_DISTANCE') := 'KMNS_EUCLIDEAN'; v_setlist('KMNS_DETAILS') := 'KMNS_DETAILS_ALL'; v_setlist('KMNS_ITERATIONS') := '3'; v_setlist('KMNS_NUM_BINS') := '10’; . . . We’ll use a k-Means algorithm to delve into which Smart Meters are clustered together, based on various attributes of the companies using them …
- 16. Clustering, Brute Force: Looking for Solar Power Super-Users (3) . . . DBMS_DATA_MINING.CREATE_MODEL2( model_name => ‘SOLAR_SUPERUSERS_CLUSTERING', mining_function => 'CLUSTERING', data_query => 'SELECT * FROM solar_superusers', set_list => v_setlist, case_id_column_name => 'SM_ID'); EXCEPTION WHEN OTHERS THEN SQLERRNUM := SQLCODE; SQLERRMSG := SQLERRM; vcReturnValue := 'Error creating new OML4SQL model: ' || SQLERRNUM || ' - ' || SQLERRMSG; DBMS_APPLICATION_INFO.SET_MODULE(NULL,NULL); DBMS_OUTPUT.PUT_LINE(vcReturnValue); END; / … and leverage the view we just created as the source for each Smart Meter’s related attributes
- 17. Clustering, Brute Force: ML Modeling Results SELECT cluster_id ,attribute_name ,operator ,numeric_value ,support ,ROUND(confidence,3) confidence FROM dm$vrsolar_superusers_clustering WHERE cluster_id IN (SELECT cluster_id FROM dm$vdSOLAR_SUPERUSERS_CLUSTERING WHERE left_child_id IS NULL AND right_child_id IS NULL) ORDER BY cluster_id ,attribute_name ,operator ,numeric_value; Now that the clustering model is complete, we can review which features were grouped together, as well as the support and confidence ratings for their placement in each of the several clusters Essentially, these are the rules that drive the clustering model to assign each subset of customers to a specific cluster
- 18. Missed It By That Much: Determining ML Model Accuracy All models are wrong, but some are useful. - George E.P Box Founder, Department of Statistics, UW- Madison The good news: All Oracle Data Mining functions offer transparency into exactly how results have been produced via accuracy determination methods specific to each mining function Mining Function Accuracy Determination Methods Clustering Centroid Classification Confusion Matrix, Lift, ROC Regression Root Mean Squared Error (RMSE), Mean Absolute Error (MAE) Time Series Mean Square Error (MSE)
- 19. Missed It By That Much: Determining ML Model Accuracy Here’s an example of using a Radar Area graph to show the top 15 business types that are the most egregious outliers in terms of solar power used (aka distance from centroid)
- 20. AutoML: Let the Database Decide! Check out the summary of all the latest AutoML enhancements! This makes it easier for “citizen data scientists” to apply the power of ML & Analytics … … the new AutoML interface makes selection of the proper algorithms a snap … … and many more new features, including Graph Studio
- 21. Classification Via AutoML: A Demonstration To illustrate how powerful a tool AutoML is, let’s run another experiment against the same data source … but this time aiming at producing models and results aimed at classification rather than regression What could possibly go wrong?
- 22. Building a Data Source for AutoML to Devour CREATE TABLE t_smartmeter_business_profiles AS SELECT sm_id ,CD.cd_minority_owned ,CD.cd_family_generations ,CD.cd_years_in_business ,CD.cd_locale_ownership ,CF.pct_profit_margin ,CF.avg_credit_score ,SM.avg_kwh_used ,SM.avg_solar_kwh ,SM.pct_solar ,SM.solar_superuser FROM t_customer_creditscoring CF . . . We’re drawing on data summarized from a Hybrid Partitioned table containing financial statistics … . . . ,t_customer_demographics CD ,(SELECT sm_id ,ROUND(AVG(smr_kwh_used),2) AS avg_kwh_used ,ROUND(AVG(smr_solar_kwh),2) AS avg_solar_kwh ,ROUND(AVG(smr_solar_kwh) / AVG(smr_kwh_used) ,2) AS pct_solar ,CASE WHEN ROUND(AVG(smr_solar_kwh) / AVG(smr_kwh_used) ,2) >= 0.15 THEN 1 ELSE 0 END AS solar_superuser FROM t_smartmeters ,t_meter_readings WHERE smr_id = sm_id GROUP BY sm_id ORDER BY sm_id) SM WHERE SM.sm_id = CF.cf_id AND SM.sm_id = CD.cd_id ORDER BY sm_id; … as well as customer demographics and solar energy usage data
- 23. Regression Experiments with AutoML (1) First, select an appropriate data source 1 AutoML automatically builds a list of potential features and their key metrics 2
- 24. Regression Experiments with AutoML (2) Review settings for prediction type, run time, model metric, and ML algorithms to apply 3 Start the experiment, choosing either speed or accuracy 4
- 25. Regression Experiments with AutoML (3) AutoML now finishes any sampling needed and moves on to feature selection 5 Next, AutoML begins building the selected models 6
- 26. Regression Experiments with AutoML (4) Model generation is complete! On to Feature Prediction Impact assessment … 7
- 27. Regression Experiments with AutoML (5) Regression(s) complete! Now let’s transform the Neural Network model into a Zeppelin notebook, with just a few mouse clicks 8
- 28. Transform an AutoML Experiment into a NoteBook (1) From the Leader Board, select one of the algorithms and click on Create Notebook 1 Name the new notebook 2
- 29. Transform an AutoML Experiment into a NoteBook (2) The new notebook is ready. Click the link to start building paragraphs and retrieving data 3 Don’t know Python? No worries! The new notebook uses OML4Py to construct paragraphs for data retrieval and modeling 4
- 30. Transform an AutoML Experiment into a NoteBook (3) Et voila! Here’s your first results from a notebook completely generated via AutoML! 5
- 31. Useful Resources and Documentation OML Algorithms “Cheat Sheet” : https://www.oracle.com/a/tech/docs/oml4sql-algorithm-cheat-sheet.pdf Oracle 21c Machine Learning Basics (including AutoML): https://docs.oracle.com/en/database/oracle/machine-learning/oml4sql/21/dmcon/machine-learning-basics.html ADW Next Generation Availability Announcement: https://www.oracle.com/news/announcement/oracle-adds-innovations-to-cloud-data-warehouse-031721.html
Editor's Notes
- #14: Support is a measure of how frequently two items occur within the same transaction. Thus, a support rating of .818 for Rule #3 means that a customer’s interest in wind turbines and expectation to renew their solar power service is likely to occur 81.8% of the time in a transaction based on this model’s results. Confidence is a measure of how often the antecedent and the consequent appear within the same transaction. Thus, for Rule #4, there is 85.3% confidence that a customer’s expectation to renew their solar power service implies they are likely to be interested in wind turbines.