Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Mbed Connect USA 2018 Workshop: Introduction to...

Mbed Connect USA 2018 Workshop: Introduction to Data Science

Takuya Kitazawa

October 22, 2018
Tweet

More Decks by Takuya Kitazawa

Other Decks in Technology

Transcript

  1. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    Introduction to Data Science Takuya Kitazawa | @takuti | [email protected] Data Science Engineer at Arm Treasure Data / Committer of Apache Hivemall
  2. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    1. Understand what Arm Treasure Data Customer Data Platform is [5mins] Data management layer of Arm Pelion IoT platform 2. Learn how machine learning and data science works [5mins] Capture characteristics of historical data, and predict unseen result 3. See every single steps of real-world data science workflow on TD [50mins] IoT-ish sample scenario based on real-life environmental data from City of Chicago 2
  3. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    Enterprise Customer Data Platform on big data infrastructure
  4. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    6 Predic:ve analy:cs on UI Data science in query language Integra:on with third-party ML toolkit For everyone who knows SQL basics For non technical people like marketers
  5. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    Data science in query language Handy, flexible way to leverage machine learning
  6. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    Difference between “Presto” and “Hive” Hive Presto Lightweight and interac:ve data access Fast ↔ Not suited for batch processing on massive data Heavy data processing task like daily batch Slow ↔ Can process massive records at once 10
  7. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    Data 3rd-party tools (e.g., visualiza6on) SQL + heavy lightweight ML with Apache Hivemall SELECT * FROM data … How to analyze your data on Treasure Data at scale 11 Schedule Treasure Workflow (a.k.a. Digdag)
  8. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    12 ‣ Scalable ML library implemented as Hive UDFs ‣ OSS project under Apache Software Foundation ‣ TD bundles Hivemall and has 3 developers (original creator + 2 core committers) https://github.com/apache/incubator-hivemall TD’s ML capability: Apache Hivemall Easy-to-use ML in SQL Scalable Runs in parallel on Hadoop ecosystem Versa:le Efficient, generic funcbons
  9. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    What ML internally does — Learning from data 13 Historical data and problem e.g., purchase log and # of sales predic:on Model Characteristics of historical data
  10. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    What ML internally does — Predicting unforeseen results 14 Model Characteristics of historical data Unforeseen data Predic:on result
  11. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    Real-world ML & data science workflow for experts 15 Problem What you want to “predict” Hypothesis & Proposal Build machine learning model Historical data Cleanse data Evaluate Deploy to produc:on
  12. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    Hivemall makes it more handy 16 Automa:cally runs in parallel on Hadoop
  13. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    Query-based simple, scalable data science workflow on TD 17 Problem What you want to “predict” Hypothesis & Proposal Build machine learning model Historical data Cleanse data Evaluate Deploy to produc:on Easily try, save, share, schedule via simple I/F in scalable manner
  14. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    Treasure Workflow: Manage highly-dependent query fragments 18 Extract Filter Interpolate Normalize … … Query Query Query Query Train data Get features Train … Query Query Query Test data Get features Predict … Accuracy Query Query Query Query https://www.digdag.io/
  15. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    Define your data science workflow in YAML format 19 _export: !include : config/general.yml td: engine: hive +prepare: call>: common/prepare_data.dig +main: +logress_train: td>: queries/logress_train.sql create_table: logress_model +compute_downsampling_rate: td>: queries/downsampling_rate.sql engine: presto store_last_results: true +logress_predict: td>: queries/logress_predict.sql create_table: prediction +evaluate: td>: queries/evaluate.sql store_last_results: true +show_accuracy: echo>: "Logloss (smaller is better): ${td.last_results.logloss}"
  16. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    Hivemall documentation http://hivemall.incubator.apache.org/userguide/ Step-by-step ML on Hivemall tutorial http://hivemall.incubator.apache.org/userguide/supervised_learning/tutorial.html Treasure Data ML workflow examples https://github.com/treasure-data/workflow-examples/tree/master/machine-learning 21
  17. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    Demo and hands-on Step-by-step guide to running data science workflow on TD
  18. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    23 https://console.treasuredata.com/users/sign_in
  19. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    24 Problem What you want to “predict” Hypothesis & Proposal Build machine learning model Historical data Cleanse data Evaluate Deploy to produc:on
  20. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    Sample scenario: City has historical energy benchmarking data 25 https://data.cityofchicago.org/Environment-Sustainable-Development/Chicago-Energy-Benchmarking/xq83-jr8c https://www.cityofchicago.org/city/en/progs/env/building-energy-benchmarking---transparency.html
  21. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    Unify variety of data on single platform 26 https://console.treasuredata.com/app/integrations/catalog Original proprietary data source
  22. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    [TASK] Upload dump benchmarking result to TD 27
  23. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    28 Problem What you want to “predict” Hypothesis & Proposal Build machine learning model Historical data Cleanse data Evaluate Deploy to produc:on
  24. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    ML problems Hivemall can solve Classifica:on - Binary: Purchase or not / Spam detecbon - Mul:-class: Tomorrow’s weather / This user’s generabon Regression - Tomorrow’s temperature / Next month’s sales / This user’s income Recommenda:on - Customers who viewed this item also viewed… 29
  25. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    ML problems Hivemall can solve Anomaly detec:on - Find excepbonally high error rate from bme series data sent by IoT device Natural language processing - Tokenize sentence and extract keywords Clustering - Grouping users based on their similaribes 30
  26. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    31 1. Predict probability of churn (i.e., binary classification) 2. Aggressively reach out “likely to churn” customers https://www.slideshare.net/TaisukeFukawa/hivemall-meetup-vol2-oisix (Japanese) Web Mobile Customer attr. Behavior on web Complaint log Source Signed-up services Actions (direct) Actions (indirect) Point Call Guide to success UI OISIX’s data Example: ML-based customer segmentation
  27. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    Our goal: Beyond rating and encouragement 
 Predict future energy consumption (electricity use) 32 https://www.cityofchicago.org/city/en/depts/mayor/supp_info/chicago-energy-benchmarking/ Chicago_Energy_Benchmarking_Beyond_Benchmarking.html
  28. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    33 Problem What you want to “predict” Hypothesis & Proposal Build machine learning model Historical data Cleanse data Evaluate Deploy to produc:on
  29. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    [TASK] Understand your data with Presto ad-hoc queries ‣ What each column means 
 https://data.cityofchicago.org/Environment-Sustainable-Development/Chicago-Energy-Benchmarking/xq83-jr8c ‣ Total number of records ‣ Benchmarking time period and frequency ‣ Distribution of different community_area and primary_property_type ‣ Max and min values in num_of_buildings and electricity_use__kbtu_ columns ‣ Missing value rate in each *_use__kbtu_ (kBtu; thousand British thermal units) column Presto aggregation functions: https://prestodb.io/docs/current/functions/aggregate.html 34
  30. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    $ pip install pandas-td 35 import pandas_td as td con = td.connect(apikey=..., endpoint=...) presto = td.create_engine('presto:mydb', con=con) hive = td.create_engine('hive:mydb', con=con) df = td.read_td('SELECT COUNT(1) FROM www_access’, presto) https://github.com/treasure-data/pandas-td
  31. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    Exploratory data analysis (EDA) with visualization 36
  32. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    37 Problem What you want to “predict” Hypothesis & Proposal Build machine learning model Historical data Cleanse data Evaluate Deploy to produc:on
  33. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    Key: Feature engineering 38 ID Community area Property type … Gross floor area Built year … Usage (kBtu) 100001 WEST TOWN Hospital … 309056 1928 … 21470037 100256 ARCHER HEIGHTS K-12 School … 447330 1990 … 35792767 … … … … … … … … 250150 NEAR NORTH SIDE Office … 335281 1912 … 24220915 Data: Historical energy consump:on at buildings Problem: Predict future electricity use 1 kWh x 3.142 = 3.142 kBtu
  34. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    Different types of features 39 Numeric objec:ve (Solubon to problem) Categorical feature Quan:ta:ve feature Timestamp / Year Need to convert e.g., elapsed years ID Community area Property type … Gross floor area Built year … Usage (kBtu) 100001 WEST TOWN Hospital … 309056 1928 … 21470037 100256 ARCHER HEIGHTS K-12 School … 447330 1990 … 35792767 … … … … … … … … 250150 NEAR NORTH SIDE Office … 335281 1912 … 24220915
  35. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    Make features “machine readable” ‣ Categorical : Set “1” or “0” to corresponding position ‣ Quantitative : Directly use it ‣ Label : “1” for purchase, “0” for non-purchase ‣ + Converting year into elapsed years as of benchmarking year 40 Community area Property type Area Building age Usage
 (kBtu) NEAR NORTH SIDE … ARCHER HEIGHTS WEST TOWN Hospital … Office (A) 
 Data year (B)
 Built year (A) - (B) 0 … 0 1 1 … 0 309055 2014 1928 86 21470037 0 … 1 0 0 … 0 447330 2014 1990 24 35792767 … … … … … … … … … … … … 1 … 0 0 0 0 1 335281 2014 1912 102 24220915
  36. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    Feature representation in Hivemall ‣ libSVM formatɹɹɹɹɹɹɹɹɹɹɹɹɹɹ 10 : 3.4, 123 : 0.5, 34567 : 0.231 ‣ can be text ɹ ɹɹɹɹɹɹɹɹɹɹ age : 86, area : 447330 ‣ -only means = 1.0ʢe.g., categoricalʣ type#office = type#office : 1.0 41 index : value or index INT BIGINT TEXT FLOAT index value ( ) index value index TEXT
  37. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    Array of quanbtabve features : select quantitative_features(array("age", "area"), 86, 447330) ["age:86.0", “area:447330"] Array of categorical features # select categorical_features(array(“commun", "type"), "NEAR NORTH", "office") [“commun#NEAR NORTH", “type#office”] * NULL is automabcally omired Hivemall internally does one-hot encoding (e.g., office → 1, 0, 0, …) Create feature vector in SQL 42 value index value index
  38. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    Feature hashing: Approximation improves scalability 43 Simplify name of quanbtabve feature and categorical feature # select feature_hashing(array("age:86", “type#office")) ["14142887:600", "10413006"] (Default upper limit: 224 + 1 = 16777217) index value index
  39. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    select id, array_concat( categorical_features( array('Chicago community area', 'Primary use of property'), community_area, primary_property_type ), quantitative_features( array('Total interior floor space', 'Building age', 'Number of buildings'), gross_floor_area___buildings__sq_ft_, data_year - year_built, num_of_buildings ) ) as features, electricity_use__kbtu_ as annual_electricity_consumption from chicago_smart_green.energy_benchmarking where electricity_use__kbtu_ is not null [TASK] Create feature vector and check output (Hive) 44
  40. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    select id, feature_hashing( add_bias( array_concat( categorical_features( array('Chicago community area', 'Primary use of property'), community_area, primary_property_type ), quantitative_features( array('Total interior floor space', 'Building age', 'Number of buildings'), gross_floor_area___buildings__sq_ft_, data_year - year_built, num_of_buildings ) ) ) ) as features, electricity_use__kbtu_ as annual_electricity_consumption from chicago_smart_green.energy_benchmarking where electricity_use__kbtu_ is not null [TASK] Advanced technique: add_bias() and feature_hashing() 45
  41. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    46 Problem What you want to “predict” Hypothesis & Proposal Build machine learning model Historical data Cleanse data Evaluate Deploy to produc:on
  42. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    Training and prediction over features-label pairs 47 Model Table Train SQL Predict SQL Usage 21M 35M … 24M How about hospital in West Town built 10 years ago? Unforeseen 0 … 0 1 1 0 0 500k 10 West Town Hospital Age ? Historically, this building consumed… features 0 … 0 1 1 … 0 309k 86 0 … 1 0 0 … 0 447k 24 … … … … … … … … … 1 … 0 0 0 … 1 335k 102
  43. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    Design features that clearly represent “characteristics” of data 48 Model Usage 21M 35M … 24M How about hospital in West Town built 10 years ago? Unforeseen 0 … 0 1 1 0 0 500k 10 West Town Hospital Age 20M Historically, this building consumed… features 0 … 0 1 1 … 0 309k 86 0 … 1 0 0 … 0 447k 24 … … … … … … … … … 1 … 0 0 0 … 1 335k 102 Similar
  44. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    Building prediction model from historical data Classification ‣ HingeLoss ‣ LogLoss (a.k.a. logistic loss) ‣ SquaredHingeLoss ‣ ModifiedHuberLoss 49 SELECT train_regressor( -- train_classifier( features, annual_electricity_consumption, '-loss squared -opt SGD -reg no -eta simple -total_steps ${total_steps}' ) as (feature, weight) FROM training Regression ‣ SquaredLoss ‣ QuantileLoss ‣ EpsilonInsensitiveLoss ‣ SquaredEpsilonInsensitiveLoss ‣ HuberLoss
  45. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    Building prediction model from historical data Optimizer ‣ SGD ‣ AdaGrad ‣ AdaDelta ‣ ADAM 50 SELECT train_classifier( -- train_regressor( features, annual_electricity_consumption, '-loss squared -opt SGD -reg no -eta simple -total_steps ${total_steps}' ) as (feature, weight) FROM training Regularization ‣ L1 ‣ L2 ‣ ElasticNet ‣ RDA ‣ Iteration with learning rate control ‣ Mini-batch training ‣ Early stopping
  46. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    52 Problem What you want to “predict” Hypothesis & Proposal Build machine learning model Historical data Cleanse data Evaluate Deploy to produc:on
  47. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    Evaluate accuracy of ML model 53 Historical data Training data Valida:on data ML model Predicbon result Measure predic:on accuracy Predicbon Actual value 38.5 30 12.1 18 25.2 20 Q. How much is this overall predic:on result good (or bad)?
  48. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    Evaluation metric example: Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) 54 Predicbon Actual value 38.5 30 12.1 18 25.2 20
  49. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    _export: !include : config.yml td: database: ${database} engine: hive +vectorize: td>: queries/vectorize.sql create_table: vectors +shuffle: td>: queries/shuffle.sql create_table: samples engine: presto +train: td>: queries/train.sql create_table: model +evaluate: td>: queries/evaluate.sql store_last_results: true +show_accuracy: echo>: "RMSE: ${td.last_results.rmse}, MAE: ${td.last_results.mae}" [TASK] Run simple regression workflow 55
  50. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    57 Problem What you want to “predict” Hypothesis & Proposal Build machine learning model Historical data Cleanse data Evaluate Deploy to produc:on
  51. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    Possible directions ‣ Design different feature vector ‣ Normalize and re-scale feature values ‣ Collect more data ‣ Join with different types of data ‣ Tweak better hyper-parameters ‣ Use other ML model 58
  52. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    Possible scenario: Chicago smart green infrastructure monitoring data https://data.cityofchicago.org/Environment-Sustainable-Development/Smart-Green-Infrastructure-Monitoring-Sensors-Hist/ggws-77ih https://github.com/BlackstoneEngineering/mbed-os-example-treasuredata-rest Each data stream captures: ‣ Temperature ‣ Wind speed and direction ‣ Rainfall ‣ Pressure ‣ Soil moisture 59 Connector chicago_smart_green.sensors_history
  53. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    60 sensors_history Environmental monitoring results from many sensors deployed in city
 community_areas (auxiliary data from CSV import) Definition of community area boundaries 
 energy_benchmarking Result of annual energy benchmarking for large buildings in city Database — chicago_smart_green Latitude, Longitude Area
  54. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    Joining different datasets on geospatial information 61 ST_Contains( ST_GeometryFromText(community_areas.the_geom), ST_Point(sensors_historical.longitude, sensors_historical.latitude) ) https://prestodb.io/docs/current/functions/geospatial.html
  55. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    62 Problem What you want to “predict” Hypothesis & Proposal Build machine learning model Historical data Cleanse data Evaluate Deploy to produc:on
  56. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    [TASK] Predict unknown NULL electricity usage 63 with null_samples as ( select id, array_concat( categorical_features( array('Chicago community area', 'Primary use of property'), community_area, primary_property_type ), quantitative_features( array('Total interior floor space', 'Building age', 'Number of buildings'), gross_floor_area___buildings__sq_ft_, data_year - year_built, num_of_buildings ) ) as features from chicago_smart_green.energy_benchmarking where electricity_use__kbtu_ is null ), features_exploded as ( select id, extract_feature(fv) as feature, extract_weight(fv) as value from null_samples t1 LATERAL VIEW explode(features) t2 as fv ) select t1.id, sum(p1.weight * t1.value) as predicted_electricity_consumption from features_exploded t1 LEFT OUTER JOIN model p1 ON (t1.feature = p1.feature) group by t1.id
  57. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    Deploy ML model to production 64 https://hivemall.incubator.apache.org/userguide/tips/rt_prediction.html Producbon env External signal Predict Vector/matrix computabon
  58. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    Predictive analytics on UI Minimal and powerful ML capabilities on unified interface
  59. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    Word-based customer tagging and categorization 68 Store customers’ browsing log from TD JavaScript SDK STEP 1 Extract keywords from each article STEP 2 Aggregate customers’ visits as td_interest_words and td_affinity_categories Society Olympic game medal president citizen rule law data cloud CDP Create audience politics law US nation equation math curry rice history Science Food, Culture td_client_id XXX-YYY-ZZZZZ td_title Today’s news td_description The Olympic game has been started … td_host www.td-news.com td_path /2017/10/01/olympic td_client_id XXX-YYY-ZZZZZ td_interest_words Olympic, baseball, game td_affinity_categories Sports, Entertainment
  60. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    td_client_id XXX-YYY-ZZZZZ td_ip 192.168.0.1 td_referrer http://google.com/… spend_time 1.5 … … td_interest_words Olympic, baseball, game td_affinity_categories Sports, Entertainment Profile set Segment What you want to predict Build predictive model Guess how to cleanse data Evaluation Japan google.com 1.5 accuracy Sufficient? Audience Unlikely Marginally Possibly Likely 12 20 3 34 40 72 58 82 93 99 78 GUESS Automatically select and transform customer attributes 1ST PASS Treasure CDP does everything for you FROM 2ND PASS You can make your predictive model better with ML experts SCORE CUSTOMERS SYNDICATE Base segment Population Predictive customer scoring 69
  61. Copyright 1995-2018 Arm Limited (or its affiliates). All rights reserved.

    Treasure Data audience suite release announcement https://blog.treasuredata.com/blog/2018/10/02/audience-suite-intuitive-actionable-customer-data-platform/ Predictive scoring documentation https://support.treasuredata.com/hc/en-us/articles/360001458407-Predicting-Customer-Behavior Developer’s tech talk slides explaining technical detail https://speakerdeck.com/takuti/machine-learning-and-natural-language-processing-on-treasure-cdp 71