Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

Meet TransmogrifAI, Open Source AutoML That Powers Einstein Predictions


Published on

Despite huge progress in machine learning over the past decade, building production-ready machine learning systems is still hard. Three years ago when we set out to build machine learning capabilities into the Salesforce platform we learned that building enterprise-scale machine learning systems is even harder.To solve the problems we encountered, we built TransmogrifAI ( (pronounced trans-mog-ri-phi), an end-to-end automated machine learning library for structured data, that is used in production today to help power our Salesforce Einstein AI platform. This talk highlights key capabilities of TransmogrifAI library and demonstrates them in action on a real-life machine learning application.

Published in: Software
  • New NFL Bankroll doubler for you. Fully Verified Proof [inside] ➤➤
    Are you sure you want to  Yes  No
    Your message goes here

Meet TransmogrifAI, Open Source AutoML That Powers Einstein Predictions

  1. 1. Meet TransmogrifAI, Open Source AutoML That Powers Einstein Predictions, @tovbinm Matthew Tovbin, Principal Engineer, Einstein
  2. 2. Forward Looking Statement Statement under the Private Securities Litigation Reform Act of 1995: This presentation may contain forward-looking statements that involve risks, uncertainties, and assumptions. If any such uncertainties materialize or if any of the assumptions proves incorrect, the results of, inc. could differ materially from the results expressed or implied by the forward-looking statements we make. All statements other than statements of historical fact could be deemed forward-looking, including any projections of product or service availability, subscriber growth, earnings, revenues, or other financial items and any statements regarding strategies or plans of management for future operations, statements of belief, any statements concerning new, planned, or upgraded services or technology developments and customer contracts or use of our services. The risks and uncertainties referred to above include – but are not limited to – risks associated with developing and delivering new functionality for our service, new products and services, our new business model, our past operating losses, possible fluctuations in our operating results and rate of growth, interruptions or delays in our Web hosting, breach of our security measures, the outcome of any litigation, risks associated with completed and any possible mergers and acquisitions, the immature market in which we operate, our relatively limited operating history, our ability to expand, retain, and motivate our employees and manage our growth, new releases of our service and successful customer deployment, our limited history reselling products, and utilization and selling to larger enterprise customers. Further information on potential factors that could affect the financial results of, inc. is included in our annual report on Form 10-K for the most recent fiscal year and in our quarterly report on Form 10-Q for the most recent fiscal quarter. These documents and others containing important disclosures are available on the SEC Filings section of the Investor Information section of our Web site. Any unreleased services or features referenced in this or other presentations, press releases or public statements are not currently available and may not be delivered on time or at all. Customers who purchase our services should make the purchase decisions based upon features that are currently available., inc. assumes no obligation and does not intend to update these forward-looking statements.
  3. 3. Multi-cloud and Multi-tenant
  4. 4. 1. Customer-specific models beat global models 2. Majority of business data is structured 3. Too many use cases, too few data scientists Machine Learning is Hard and Even Harder for the Enterprise Lessons our Data Scientists Learned while Building Einstein
  5. 5. 1. Customer-specific Models Beat Global Models ● Customers care about data privacy ● Every customer’s data is different Enterprise Machine Learning
  6. 6. 2. Majority of Business Data is Structured
  7. 7. Data Prep Feature Engineering Feature Selection Model Training Model The standard approach to building an ML model 3. Too Many Use Cases, Too Few Data Scientists
  8. 8. ML is exponentially harder in the Enterprise with many, customer-specific models 3. Too Many Use Cases, Too Few Data Scientists Data Prep Feat. Eng Feat. Selection Model Training Model Data Prep Feat. Eng Feat. Selection Model Training Model Data Prep Feat. Eng Feat. Selection Model Training Model Data Prep Feat. Eng Feat. Selection Model Training Model Data Prep Feat. Eng Feat. Selection Model Training Model Data Prep Feat. Eng Feat. Selection Model Training Model Data Prep Feat. Eng Feat. Selection Model Training Model Data Prep Feat. Eng Feat. Selection Model Training Model
  9. 9. TransmogrifAI Introducing TransmogrifAI Customer specific models Structured, transactional data Data science at scale + + Automated Machine Learning for Structured Data
  10. 10. ● Automated feature engineering, feature selection & model selection ● ML abstractions that improve developer productivity & collaboration ● Model explainability to improve debuggability and transparency >90% accuracy with 100x reduction in time Introducing TransmogrifAI Automated Machine Learning for Structured Data
  11. 11. Transform in a surprising or magical manner What’s in a name? transmogrify
  12. 12. 5B+ predictions per day Einstein Platform Compute Orchestration Data Store Model Lifecycle Management Data Science Experience Configuration Services Infrastructure Metrics Health Monitoring ETL/GDPR/ Data Processing DL TransmogrifAI Machine Learning The AutoML Engine in the Einstein Platform Lead Scoring Engagement ScoringCase Classification Prediction Builder ...
  13. 13. Einstein Prediction Builder • Product: Point. Click. Predict. • Engineering: any customer can create any number of ML applications on any data?! Impossible!
  14. 14. Under the Hood ● Automated Feature Engineering ● Automated Feature Selection ● Automated Model Selection
  15. 15. Automatic Feature Engineering
  16. 16. Type Hierarchy For Machine Learning FeatureType OPNumeric OPCollection OPSetOPList NonNullableText Email Base64 Phone ID URL ComboBox PickList TextArea OPVector OPMap BinaryMap IntegralMap DateList DateTimeList Integral Real Binary Percent Currency Date DateTime MultiPickList TextMap TextListCity Street Country PostalCode Location State Geolocation StateMap SingleResponse RealNN Categorical MultiResponse Legend: bold - abstract type, regular - concrete type, italic - trait, solid line - inheritance, dashed line - trait mixin ... RealMap Prediction
  17. 17. Automatic Feature Engineering transmogrify() Lat LonSubjectPhoneEmail Age Age [0-15] Age [15-35] Age [>35] Email Is Spammy Top Email Domains Country Code Phone Is Valid Top TF-IDF Terms City, State Feature Vector
  18. 18. Feature 34,200.03 14.001.02 22,430.11 47,895.66 Feature Null Indicator 34,200.03 0 14.001.02 0 16,045.21 1 22,430.11 0 16,045.21 1 47,895.66 0 Numeric – Imputation and Null value tracking
  19. 19. Categorical: One Hot Encoding
  20. 20. Text: TF-IDF
  21. 21. Temporal: Circular Statistics Circular distributions are those that have no true zero. Great for temporal features and deals with seasonality: ● Hours of the Day ● Weeks on the Month ● Months of the Year
  22. 22. Numeric Categorical SpatialTemporal Reverse Geocoding Nearest POI Text Time difference Circular Statistics Time extraction (day, week, month, year) Language Detection Language-wise Tokenization Hash Encoding Tf-Idf Word2Vec Name Entity Resolution Smart Categorical Imputation Track null value One Hot Encoding Dynamic Top K pivot Imputation Track null value Scaling - zNormalize, log, linear Smart Binning Automatic Feature Engineering
  23. 23. Automatic Feature Selection
  24. 24. Problems with doing Machine Learning on Enterprise Data 1. Hindsight Bias 2. Field Usage Changes 3. Bulk Uploads 4. Field Type Abuse 5. More...
  25. 25. Lead Before Conversion Lead At Conversion Problem #1 – Hindsight Bias (aka Label Leakage)
  26. 26. In layman terms, it is like Marty McFly traveling to the future, getting his hands on the Sports Almanac, and using it to bet on the games of the present.
  27. 27. Problem #2 – Field Usage Changes Over Time
  28. 28. Problem #3 – Bulk Upload by Business Workflow A business process updated records having different distribution - biased towards negative outcome
  29. 29. The quick, brown fox jumps over a lazy dog. DJs flock by when MTV ax quiz prog. Junk MTV quiz graced by fox whelps. Bawds jog, flick quartz, vex nymphs. Waltz, bad nymph, for quick jigs vex! Fox nymphs grab quick-jived waltz. Brick quiz whangs jumpy veldt fox. Bright vixens jump; dozy fowl quack. Quick wafting zephyrs vex bold Jim. Quick zephyrs blow, vexing daft Jim. Sex-charged fop blew my junk TV quiz. How quickly daft jumping zebras vex. Two driven jocks help fax my big quiz. Quick, Baz, get my woven flax jodhpurs! "Now fax quiz Jack!" my brave ghost pled. Five quacking zephyrs jolt my wax bed. Flummoxed by job, kvetching W. zaps Iraq. Cozy sphinx waves quart jug of bad milk. A very bad quack might jinx zippy fowls. Few quips galvanized the mock jury box. Quick brown dogs jump over the lazy fox. The jay, pig, fox, zebra, and my wolves quack! Blowzy red vixens fight for a quick jump. Joaquin Phoenix was gazed by MTV for luck. A wizard’s job is to vex chumps quickly in fog. Watch "Jeopardy!", Alex Trebek's fun TV quiz game. Woven silk pyjamas exchanged for blue quartz. Brawny gods just Typical Text Feature ‘Last Open Stage’ Text Feature align answer collect contracting negotiate opportunity won qualify qualify/align Problem #4 – Feature types abused
  30. 30. outcome/label Opportunity Won value of this feature is a leaker Problem #4 – Feature types abused
  31. 31. ● Analyze every feature and output descriptive statistics ○ Mean ○ Min ○ Max ○ Variance ○ Number of Nulls ● Ensure Features have acceptable ranges Automatic Feature Selection
  32. 32. ● Analyse each feature correlation to the label, who has the most and least predictive power? ● Drop features with low predictive power Automatic Feature Selection
  33. 33. Auto Bucketize training vs scoring Feature Lineage
  34. 34. Need to know the true label to evaluate the model ● Usually do a random train/holdout split on the labeled data and use cross-validation on training set Evaluating Models Training set Holdout set
  35. 35. ● Time-based evaluation dataset is the true test of how well a model is performing ○ Wait for existing (or new) records to have their label determined ○ Predict from older state of that record and compare to the true label ● Biggest problem is usually waiting for enough data to be available ● We can also switch over to constructing the model from the true event sequence rather than a snapshot Evaluating Models
  36. 36. What does label leakage look like?
  37. 37. What does label leakage look like?
  38. 38. Leakers removed by AutoML: 73 Leakers removed by data scientist hand tuning: 42 Department mkto_si__Last_Interesting_Moment__c Description OtherPostalCode et4ae5__Mobile_Country_Code__c Title mkto2__Acquisition_Program_Id__c JigsawContactId ReportsToId OtherCity pi__last_activity__c MailingLongitude pi__first_activity__c AssistantPhone HomePhone Fax OtherStreet Partner_Last_Name__c mkto_si__Last_Interesting_Moment_Desc__c mkto2__Acquisition_Program__c Jigsaw Company__c OtherLongitude AssistantName Salutation OtherLatitude Purchase_Motivation__c Secondary_Email__c TimetoPurchase__c mkto_si__Last_Interesting_Moment_Source__c MailingGeocodeAccuracy MailingLatitude pi__created_date__c CommentCapture__c Preferred_Communication_Method__c TopPriorityValue__c mkto_si__Last_Interesting_Moment_Type__c OtherState TopPriorityProcess__c OtherCountry MasterRecordId OtherGeocodeAccuracy TopPriorityProduct__c emailbounceddate lastcurequestdate lastcuupdatedate lastreferenceddate lastvieweddate mkto2__acquisition_date__c mkto_si__hidedate__c pi__grade__c pi__notes__c pi__utm_content__c account_link_easy_closets__c csat_survey_completed_date__c csat_survey_net_promoter_score__c csat_survey_results_link__c birthdate mkto_si__last_interesting_moment_date__c pi__campaign__c pi__comments__c pi__first_search_term__c pi__first_search_type__c pi__first_touch_url__c pi__score__c pi__url__c pi__utm_campaign__c pi__utm_medium__c pi__utm_source__c historical_lead_score__c pi__utm_term__c first_activity_timestamp__c predicted_likelihood_to_purchase_2__c best_time_to_call_date__ c total_lead_score__c csat_customer_service_s urvey_disallowed__c referral_credit_applied__c referral_days_til_purchas e__c predicted_likelihood_to_p urchase__c createdbyid createddate lastactivitydate lastmodifieddate last_activity_date__c systemmodstamp AutoML vs Hand Tuned – Showdown
  39. 39. Live Prediction Results AutoML vs Hand Tuned – Showdown
  40. 40. Automated Model Selection
  41. 41. Automated Model Selection ● Many hyperparameters for each algorithm ● Automated Hyperparameter tuning ○ Faster model creation with improved metrics ○ Search algorithms to find the optimal hyperparameters, e.g grid search, random search Grid Search Bayesian SearchRandom Search
  42. 42. Random Forests Decision Trees Logistic Regression w/ ElasticNet Regularization Naive Bayes Gradient Boosted trees Decision Trees Random Forests Linear Regression w/ ElasticNet Regularization Random Forests Decision Trees Multinomial Logistic Regression w/ ElasticNet Naive Bayes Compete Algorithms RMSE AccuracyAuROC Regression Binary Classification Multi-Class Classification Automated Model Selection
  43. 43. Different Permutation of Thresholds Leads to Different Results
  44. 44. Demo Image credit: Wikipedia
  45. 45. How well does it work? • TransmogrifAI empowers: • Predictive Journeys • Lead Scoring • Prediction Builder • Case Classification • Most of the models deployed in production are completely hands free • Serves 3B+ 5B+ predictions per day
  46. 46. Where do WE go next? • Deeper model & score insights – LOCO, LIME • Hyper parameter search strategies – Bayesian, Bandit-based • Feature engineering – text embeddings, model specific • Model portability • Enable more applications – recommenders, unsupervised learning • Perf tuning, bug fixes, docs, examples • <Your requirements / feedback>
  47. 47. Where do YOU go next? • Read the blog post - • Try it out - • Reach out and contribute - • Student? Apply to Google Summer of Code (GSoC) 2019 to work with us! • Feeling creative? We need a logo.
  48. 48. Questions?