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

Time Series Data Prediction for Service Ops

Time Series Data Prediction for Service Ops

LINE DEVDAY 2021

November 11, 2021
Tweet

More Decks by LINE DEVDAY 2021

Other Decks in Technology

Transcript

  1. Agenda - Project Background & Members
    - Overview of In-House Time Series Prediction
    Tool
    - Achievements and Future Prospects

    View full-size slide

  2. Project Members
    Lu Yan
    Data Scientist
    Terry So
    Machine Learning Engineer
    Yuan Yifan
    Machine Learning Engineer

    View full-size slide

  3. Typical use case for user
    inquiry form
    à User buys new iPhone
    à Has problem transferring
    LINE account to the new
    device
    à Contact LINE via an
    inquiry form
    Project Background

    View full-size slide

  4. Project Background
    a Typical Use Case
    LINE Fukuoka Data Scientist Team + Value Management Center
    < Inquiry Count Forecasting>
    How many users’ inquiries do we need to treat
    for each LINE service every month? Every day? Every hour?

    View full-size slide

  5. Project Background
    Motivation for Developing a Forecasting Tool
    - Dozens of predictions required every month for different departments
    - High project communication cost
    - Limited number of Data Scientists
    - Relatively similar data structure
    - Many repetitive operations of data pre-treatment and model training

    View full-size slide

  6. Project Background
    - Develop a tool for both Data Scientists and non-Data Scientist users from
    LINE biz/operational service department
    - For non-DS(non-Data Scientist) users
    - Let users do the forecast by themselves
    - For DS(Data Scientist)
    - Automate repetitive steps
    Motivation for Developing a Forecasting Tool

    View full-size slide

  7. For non-DS For DS
    Project Background
    Tool Design

    View full-size slide

  8. Agenda - Project Background & Members
    - Overview of In-House Time Series Prediction
    Tool
    - Achievements and Future Prospects

    View full-size slide

  9. Main Components of the Tool
    Add Dataset
    Train the
    model
    Forecast
    Add Project
    Overview of In-House Time Series Auto
    Prediction Tool

    View full-size slide

  10. Main Components of the Tool - Project
    Overview of Tool – for non-DS User

    View full-size slide

  11. Main Components of the Tool
    Add Dataset
    Train the
    model
    Forecast
    Add Project
    Overview of In-House Time Series Auto
    Prediction Tool

    View full-size slide

  12. Main Components of the Tool – Dataset
    Overview of Tool – for non-DS User
    - Supports different kind of data sources (CSV,
    MySQL, Presto, Spark SQL etc.)
    - Connects to internal DB
    - Provides anomaly detection and imputation
    function

    View full-size slide

  13. Main Components of the Tool – Dataset
    Overview of Tool – for non-DS User

    View full-size slide

  14. Main Components of the Tool
    Add Dataset
    Train the
    model
    Forecast
    Add Project
    Overview of In-House Time Series Auto
    Prediction Tool

    View full-size slide

  15. Main Components of the Tool - Predictor
    Overview of Tool – for non-DS User
    - Different types of models
    (Prophet, LSTM, ARIMA,
    Temporal Fusion Transformers
    etc..)
    - Hyperparameter Tuning
    - Model Explanation
    - Model Evaluation
    (RMSE/MAPE/WAPE etc..)

    View full-size slide

  16. Main Components of the Tool
    Add Dataset
    Train the
    model
    Forecast
    Add Project
    Overview of In-House Time Series Auto
    Prediction Tool

    View full-size slide

  17. Overview of Tool – for non-DS User
    Main Components of the Tool - Forecast

    View full-size slide

  18. - Runs Python client
    - Supports batch jobs
    - Provides advanced functions
    - Keeps adding new features
    Overview of Tool – for Data Scientist

    View full-size slide

  19. Overview of Tool – for Data Scientist
    Batch Jobs

    View full-size slide

  20. Overview of Tool – for Data Scientist
    Advanced Functions Examples – Hyperparameter Tuner

    View full-size slide

  21. Overview of Tool – for Data Scientist
    Advanced Functions Examples – Model Explanation

    View full-size slide

  22. Overview of Tool – for Data Scientist
    Advanced Functions Examples – Model Explanation

    View full-size slide

  23. - Support usage of standard SQL queries to run the model and to do the prediction
    Overview of Tool – for Data Scientist
    Keep Adding New Features – SQL-Like Syntax

    View full-size slide

  24. Overview of Tool – for Data Scientist
    Keep Adding New Feature - Show Past Predictions

    View full-size slide

  25. Overview of Tool
    System Architecture

    View full-size slide

  26. Overview of Tool
    System Architecture
    Django Framework
    – Web API

    View full-size slide

  27. Overview of Tool
    System Architecture
    Celery – Distributed Task Queue

    View full-size slide

  28. Overview of Tool
    System Architecture
    Elasticsearch
    - Log System

    View full-size slide

  29. Agenda - Project Background & Members
    - Overview of In-House Time Series Prediction
    Tool
    - Achievements and Future Prospects

    View full-size slide

  30. Achievements and Future Prospects
    Quantity of prediction tasks
    + High project communication cost
    VS Limited number of Data Scientists
    Repetitive operations of data pre-
    treatment and model training
    DS can take advantage of the API &
    advanced functions and can run
    batch jobs easily
    Non-DS Users can do predictions
    by themselves with the UI and
    functionalities developed
    Use Case with Value Management Center

    View full-size slide

  31. Achievements and Future Prospects
    Use Case with Value Management Center
    - Use case – Inquiry Count
    Forecasting
    - After the usage of this tool, the
    work time has been reduced by
    83%
    - The accuracy of prediction is
    higher and more stable,
    regardless of seasonality and
    personnel change.
    Source :https://linefukuoka.blog.jp/archives/20210630_01_news_plan_and_op.html

    View full-size slide

  32. Achievements and Future Prospects
    - Open to more business departments
    - Improve UI
    - Support more types of data visualization
    - Allow users to select weather/holiday common features
    - Support multi-horizon forecast
    Future Prospects

    View full-size slide