O slideshow foi denunciado.
Utilizamos seu perfil e dados de atividades no LinkedIn para personalizar e exibir anúncios mais relevantes. Altere suas preferências de anúncios quando desejar.

Come costruire servizi di Forecasting sfruttando algoritmi di ML e deep learning.

1.755 visualizações

Publicada em

Il Forecasting è un processo importante per tantissime aziende e viene utilizzato in vari ambiti per cercare di prevedere in modo accurato la crescita e distribuzione di un prodotto, l’utilizzo delle risorse necessarie nelle linee produttive, presentazioni finanziarie e tanto altro. Amazon utilizza delle tecniche avanzate di forecasting, in parte questi servizi sono stati messi a disposizione di tutti i clienti AWS.
In questa sessione illustreremo come pre-processare i dati che contengono una componente temporale e successivamente utilizzare un algoritmo che a partire dal tipo di dato analizzato produce un forecasting accurato.

  • Entre para ver os comentários

Come costruire servizi di Forecasting sfruttando algoritmi di ML e deep learning.

  1. 1. 1© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved | How to build Forecasting using ML/DL algorithms Iona Ekonomi – Senior Solutions Architect
  2. 2. 2© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved | Introduction Forecasting Techniques Amazon Forecasting Amazon SageMaker simplify what you need to know to deploy the Anaplan on AWS solution quickly Agenda
  3. 3. 3© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved | The AWS ML Stack Broadest and most complete set of Machine Learning capabilities VISION SPEECH TEXT SEARCH NEW CHATBOTS PERSONALIZATION FORECASTING FRAUD DEVELOPMENT CONTACT CENTERS Amazon SageMaker Ground Truth Augmented AI SageMaker Neo Built-in algorithms SageMaker Notebooks SageMaker Experiments Model tuning SageMaker Debugger SageMaker Autopilot Model hosting SageMaker Model Monitor Deep Learning AMIs & Containers GPUs & CPUs Elastic Inference Inferentia FPGA Amazon Rekognition Amazon Polly Amazon Transcribe +Medical Amazon Comprehend +Medical Amazon Translate Amazon Lex Amazon Personalize Amazon Forecast Amazon Fraud Detector Amazon CodeGuru AI SERVICES ML SERVICES ML FRAMEWORKS & INFRASTRUCTURE Amazon Textract Amazon Kendra Contact Lens For Amazon Connect SageMaker Studio IDE NEW NEW NEW
  4. 4. 4© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved | 1995 2000 2007 2010 2015 2019 Forecasting at Amazon.com Using machine learning to solve complex forecasting problems Use of Machine Learning High price variability Slow moving productsRegional vs national demand New products Highly seasonal products Traditional statistical methods Use of deep learning 15x Improvement in accuracy
  5. 5. 5© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved |
  6. 6. 6© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved | Private forecasting API Amazon Forecast The technology that powers the world’s largest ecommerce business Get started with the console or API Point Amazon Forecast to your data stored in Amazon Simple Storage Service (Amazon S3) Automatically train your custom ML model Let Amazon Forecast auto select the best one for your data through AutoML Generate accurate forecasts Retrieve forecasts through the console or private API Historical data Related data Sales, call volume, inventory, resource demand Price, promotions, weather data, custom events Item metadata Color, city, country, category, author, album name Built-in dataset (holiday, weekends) Amazon Forecast
  7. 7. 7© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved | Customized forecasting API Inspect data Identify features Select most accurate model from multiple algorithms Select Hyper- parameters Host models Load data Train models using multiple algorithms Optimize models Amazon Forecast Behind the scenes Fully managed by Amazon Forecast Historical data Related data sales, call volume, inventory, resource demand. Price, promotions, weather data, custom events Item metadata Color, city, country, category, author, album name Create dataset Create predictor (train, inference, metrics) Create Forecast Create Forecast export Query Forecast Amazon Forecast
  8. 8. 8© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved | Use of historical data to predict future values Target time- series dataset The primary variable to predict with its historical values (demand, sales) Datasets used for forecasting Use of related attributes and categorical data Item metadata (non-time- varying) Categorical data that provide more context about items (color, city, channel) Use of known time- varying data specific to your business Related time- series dataset Time-varying related features that may impact the target value (price, promotion, weather) Amazon Forecast
  9. 9. 9© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved | Predictor • Custom model trained on your data. • A forecast horizon – how far you want to predicate also called the prediction length. • the maximum forecast horizon is the lesser of 500 time-steps or 1/3 of the TARGET_TIME_SERIES dataset length. • Evaluation parameters – How to split a dataset into training and test datasets using backtest • Then either you chose the algorithm manually or make it Auto where AWS will try all algorithms and choice the best one. • AutoML optimizes the average of the weighted P10, P50 and P90 quantile losses, and returns the algorithm with the lowest value.
  10. 10. 10© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved | Forecast • It’s your model deployed on the production on somewhere on AWS cloud and is fully managed by AWS to match your demand. And now all you need to call this end point to get the results using Query Forecast. • Call the CreateForecast operation to create a forecast. • During forecast creation, Amazon Forecast trains a model on the entire dataset before hosting the model and doing inference. • This operation creates a forecast for every item (item_id) in the dataset group that was used to train the predictor. • After a forecast is created, you can query the forecast or export it to your Amazon Simple Storage Service (Amazon S3) bucket. Private forecasting API
  11. 11. 11© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved | Visualize the distribution of forecasted values View probabilistic forecasts at any quantile in the console Retrieve forecasts through your private API Export forecasts to .csv Amazon Forecast
  12. 12. 12© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved | Handles tricky forecasting scenarios Missing values Cold start (new product introduction) Irregular seasonality Product discontinuation Highly spiky data Sensitivity analysis (future price change) Amazon Forecast
  13. 13. 13© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved | Amazon Forecast Algorithms +
  14. 14. 14© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved | ARIMA Auto-regressive integrated moving average AMAZON FORECAST - ARIMA • The last step move from ARMA to ARIMA is differencing step called integrate ARIMA(p,d,q). • So we do it on two stages • First apply differencing (order d) • Then ARMA (p,q)
  15. 15. 15© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved | AMAZON FORECAST - ARIMA Linear regression • Linear regression attempts to model the relationship between two variables by fitting a linear equation to observed data. • WE can use linear regression in forecast y=mx+b • The slope of the line is m (coefficient), and b is the intercept (the value of y when x = 0). Capture the trend and seasonality
  16. 16. 16© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved | AMAZON FORECAST - ARIMA Autocorrelation • Some time called serial correlation. • Find the correlation between the series and its past value to improve the forecast. • Correlation between pairs of values at a certain lag. • Lag-1 autocorrelation : Yt and Yt-1 • Lag-2 autocorrelation : Yt and Yt-2
  17. 17. 17© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved | Autoregression (AR Model) • Capture autocorrelation in a series in an regression type model and use it to improve short-term forecasts key concept is Order. AR(p) ARMA(p,q) Only work for short term Forecast Autoregression Moving Average (ARMA) • It require Stationarity -no trend/s seasonality • So we can have apply on two stage 1. capture trend using regression. 2. apply AR model to capture autocorrelation and next forecast error [Moving Average] 3. Combine the two to get the improve forecast AMAZON FORECAST - ARIMA
  18. 18. 18© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved | AMAZON FORECAST - ETS ETS (ExponenTial Smoothing) Error Trend Seasonality Statistical algorithm that uses exponential smoothing Exponential smoothing forecasting : prediction is a weighted sum of past observations, but the model explicitly uses an exponentially decreasing weight for past observations.
  19. 19. 19© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved | AMAZON FORECAST - ETS Smoothing create an approximating function that attempts to capture important patterns in the data, while leaving out noise or other fine-scale structures/rapid phenomena
  20. 20. 20© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved | AMAZON FORECAST – ETS Holt’s exponential smoothing Smoothing using differencing : • help in the case of dataset has trends and seasonality • Differencing means taking the difference between two values of the series • lag :- means how far apart these two value are for example lag = 1 mean y(t) - y(t-1) which help to remove trend. • lag-M differencing y(t) - y(t-m) useful for removing seasonality with M seasons. Holt’s exponential smoothing (double exponential smoothing ): • Fore series with trend but no seasonality. • 𝐹!"# = 𝐿! + K 𝑇! ( T is the trend ) • 𝑇! = 𝛽 (𝐿! − 𝐿!$% ) + (1- 𝛽) 𝑇!$% • 0 ≤ 𝛽 ≤ 1 à how fast we update the trend • 𝛽à is the trend constant
  21. 21. 21© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved | AMAZON FORECAST – ETS Winter’s exponential smoothing (triple exponential smoothing ) • Fore series with trend & seasonality • 𝐹!"# = 𝐿! + K 𝑇! + 𝑆!"#$& • 𝑆! = γ (𝐿! − 𝐿!$% ) + (1- 𝛾) 𝑆!$% • 0 ≤ 𝛾 ≤ 1 à how fast we update the seasonality • Forecast = most recent estimated level + trend + seasonality.
  22. 22. 22© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved | AMAZON FORECAST – NPTS NPTS Non-parametric time series Jan 06 2014 Apr 07 2014 Jul 07 2014 Oct 06 2014 Jan 05 2015 Apr 06 2015 Jul 06 2015 Oct 05 2015 Jan 04 2016 Apr 04 2016 0246810 A Typical Time Series in Large Inventories
  23. 23. 23© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved | AMAZON FORECAST – NPTS Parametric • Fixed number of parameters • computationally faster, but makes stronger assumptions about the data. • A common example of a parametric algorithm is linear regression. • we try to find y=mx+b then we though the data away and use the equation in the future to find y Non-Parametric • uses a flexible number of parameters and grows as it learns from more data • computationally slower • example is K-nearest neighbour and kernel regression • we keep the data and we always come back and consult the data to find the right predication
  24. 24. 24© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved | AMAZON FORECAST – Prophet Prophet Additive regression model with Gaussian likelihood Can find trend, seasonality, cyclical, and holiday effects
  25. 25. 25© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved | • Structural time series model develop by Facebook became opensource in Feb 2017. • Use a very flexible regression model (somewhat like curve- fitting) • Builds model by finding a best smooth line which as sum of • Overall growth trend • Yearly seasonality • Weekly seasonality • Holiday effects – X’mas, New Year etc. AMAZON FORECAST – Prophet
  26. 26. 26© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved | AMAZON FORECAST – DEEPAR+ DeepAR+ Supervised learning algorithm based on autoregressive RNNs that can produce both point and probabilistic forecasts . Based on LSTM Networks Global model that can use related time series and attributes
  27. 27. 27© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved | Forecast Autoregressive history Covariates Neural networks are good at leveraging long history to learn its influence on future points, and they can handle high-dimensionality in the inputs (that is, they can handle many related-items). AMAZON FORECAST – DEEPAR+
  28. 28. 28© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved | AMAZON FORECAST – DEEPAR+ Missing data
  29. 29. 29© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved | • The DeepAR+ forecasting algorithm has been used internally in Amazon for mission-critical decisions •Classical forecasting techniques such as ARIMA and ETS fit one model to an individual time series. However, in many situations, a set of related time series have been or can be collected. •DeepAR+ can train a model over such a set of related time series for additional insights and increased predictive power •Requires minimal feature engineering and can produce forecasts that are either point (amount sold was X) or probabilistic (amount sold was between X and Y with Z probability). AMAZON FORECAST – DEEPAR+
  30. 30. 30© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved | AMAZON FORECAST – DEEPAR+ Feature Engineering, Custom Feature
  31. 31. 31© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved | Amazon SageMaker Fast & accurate data labeling Built-in, high- performance algorithms & notebooks Build 1 One-click training and tuning Train Model optimization 2 Deploy 3 Fully managed hosting with auto-scaling and elastic inference One-click deployment Build, train, and deploy ML models at scale
  32. 32. 32© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved | Amazon SageMaker Notebooks • Jupyter notebooks • Support JupyterLab • Multiple built-in kernels • Install external libraries • Install external kernels • Integrate with Git • Sample notebooks
  33. 33. 33© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved | One click in console Using API/SDK - OR - Launch training Amazon SageMaker training
  34. 34. 34© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved | Amazon SageMaker built-in algorithms Supported frameworks AWS Marketplace algorithms Model Data Data Data Data Orchestration Built-in algorithms AmazonSageMaker Model Model Model Orchestration AmazonSageMaker Custom script Algorithms or models Custom script on supported frameworks BYO algorithm and framework 17 built-in high- performance algorithms Supported frameworks: Apache MXNet, TensorFlow, Scikit-learn, PyTorch, Chainer Docker containers with your own algorithms and frameworks Third-party algorithms and models Supported frameworks Orchestration AmazonSageMaker Custom script and custom framework Orchestration AmazonSageMaker Amazon SageMaker training
  35. 35. 35© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved |
  36. 36. 36© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved |
  37. 37. 37© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved |
  38. 38. 38© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved |
  39. 39. 39© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved |
  40. 40. 40© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved |
  41. 41. 41© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved |
  42. 42. 42© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved |
  43. 43. 43© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved |
  44. 44. 44© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved |
  45. 45. 45© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved | AWS Europe (Milan) Region On April, 28th AWS expanded its global footprint with the opening of the AWS Infrastructure Region in Italy. The new Region AWS Europe (Milano) brings advanced cloud technologies that enable opportunities for innovation, entrepreneurship, and digital transformation. For additional information about services and characteristics of an AWS Region, you can check the website: aws.amazon.com/local/italy/milan/
  46. 46. 46© 2020 Amazon Web Services, Inc. or its affiliates. All rights reserved | Thanks!

×