Presented By
O’REILLY + INTEL AI

English中文
PUT AI TO WORK
June 18-21, 2019
Beijing, CN

基于深度学习的时间序列预测 (Deep Learning for Time Series Forecasting)

此演讲使用中文 (This will be presented in Chinese)

Yijing Chen (Microsoft), Dmitry Pechyoni (Microsoft), Angus Taylor (Microsoft), Vanja Paunic (Microsoft), Henry Zeng (Microsoft)
09:0012:30 Wednesday, June 19, 2019
模型与方法 (Models and Methods)
Location: 报告厅(Auditorium)

必要预备知识 (Prerequisite Knowledge)

- Understand the basic concepts of machine learning - Understand the basic concepts of neural networks - Have experience of building machine learning models

您将学到什么 (What you'll learn)

By attending the tutorial, the audience will: 1. Understand the basics of convolutional and recurrent neural networks and their advanced architectures that are effective in time series forecasting 2. Know when to use deep learning models instead of traditional time series models in time series forecasting 3. Know a number of techniques and tricks that are important for building successful deep learning models for time series forecasting 4. Get access to the source code of examples of how to train and tune deep learning models for time series forecasting using Keras

描述 (Description)

Time series forecasting is one of the most important topics in data science. Almost every business needs to predict the future in order to make better decisions and allocate resources more effectively. Examples of time series forecasting use cases are: financial forecasting, product sales forecasting, web traffic forecasting, energy demand forecasting for buildings and data centers and many more. However, most existing forecasting solutions use traditional time series and machine learning models. For complex forecasting problems, data scientists need to know how to leverage advanced techniques to generate more accurate forecasts.

Deep neural networks have achieved a lot of success for many applications. In particular, Recurrent neural networks (RNNs) are frequently used in text, speech and video analysis, being designed for processing sequential data. Additionally, Convolutional neural networks (CNNs) have achieved state-of-the-art performance on many computer vision tasks. These methods have only recently been applied to the task of time series forecasting but typically deliver strong performance. In this tutorial, we describe the basic concepts for building such models and show how and when to apply them to time series forecasting.

The tutorial will start with a quick overview of time series forecasting and neural network to provide the audience with a clear background on the kind of problems that we aim to solve.

Then we will give a comprehensive introduction starting from convolutional neural networks (CNNs) to recurrent neural networks (RNNs) such as LSTM and GRU. Each introduction will end with a hands-on exercise to help the audience solid the knowledge foundation before we move to advanced topics.
 
In the second part of the tutorial we will advance to how to apply Encoder-decoder RNN architecture to time series forecasting, followed by how to build state of art hybrid model with both CNNs and RNNs. Then we will show audience how to do hyper-perimeter tuning for such advanced model in real application.

Below is a summary of topics that we will cover in this tutorial:
· Tutorial Introduction
o Tutorial goals
o Tutorial agenda
o Target audience
o Why neural networks model for time series forecasting

· Knowledge Recap
o Time series recap
○ Time Series & Time Series Forecasting
○ Why is Time Series Forecasting important?
○ Questions to ask before building forecast model
o Feedforward Neural Network recap
○ Perceptron & multilayer perceptron​
○ Backpropagation​

· Introduction to convolutional neural networks (CNNs)
o Introduction to CNNs
o How CNNs are trained
o Hands-on exercise: How to apply CNNs to time series forecasting

· Introduction to recurrent neural networks (RNNs)
○ What are RNNs? ​
○ How RNNs are trained: Backpropagation through time (BPTT)​
○ Vanilla RNN and its gradient problems​
○ Other RNN units​
○ LSTM​
○ GRU​
o RNN stacking
o Hands-on exercise: How to apply RNNs to time series forecasting

·  Encoder-decoder RNN model
o What is encoder-decoder RNN architecture
o How to apply encode-decode RNN to multi-step time series forecasting

· Build state of art models for time series forecasting
o Hybrid CNN + RNN
o Hybrid traditional time series + RNN
o Hybrid RNN + feedforward
o Hands-on exercise

· Hyper-parameter tuning
o How to do hyper-parameter tuning
o Hands-on exercise

· Conclusion & Key takeaways

Photo of Yijing Chen

Yijing Chen

Microsoft

Yijing Chen is a senior data scientist in the Cloud AI Group at Microsoft, where she works with external customers in areas such as energy demand forecast, user mobile behavioral analysis, retail demand forecast, energy theft detection, product pricing, and medical claim denial prediction as well as on other projects using various machine learning methods. Yijing holds an MA in statistics from Harvard University.

Photo of Dmitry Pechyoni

Dmitry Pechyoni

Microsoft

Dmitry Pechyoni is a senior data scientist in the Cloud AI Group at Microsoft, where he works on building end-to-end data science solutions in various domains, including retail, energy management, and predictive maintenance. Previously, he built machine learning models for display advertising Akamai and MediaMath. Dmitry holds a PhD in theoretical machine learning from the Technion – Israel Institute of Technology.

Photo of Angus Taylor

Angus Taylor

Microsoft

Angus Taylor is a data scientist in the Cloud AI Group at Microsoft, where he builds data science solutions for external customers in the retail, energy, engineering, and package distribution sectors. He holds an MSc in AI from the University of Edinburgh.

Photo of Vanja Paunic

Vanja Paunic

Microsoft

Vanja Paunic is a data scientist in the Algorithms and Data Science Group at Microsoft London. She works on building machine learning solutions with external companies utilizing Microsoft’s AI Cloud Platform. She holds a PhD in computer science with a focus on data mining in the biomedical domain from the University of Minnesota.

Photo of Henry Zeng

Henry Zeng

Microsoft

Henry Zeng is a principal program manager in the Cloud AI Group at Microsoft, where he works with engineering team, partners and customers to ensure the success of ML platform. He has been in AI and data area for more than 10 years from database, NoSQL, Hadoop ecosystem, machine learning to deep learning. Prior to this role, he was the lead AI solution architect in Microsoft China working with partners and customer to land AI solutions in manufactory, retail, education and public service etc with Microsoft AI offerings. Henry holds a MS in computer science from Wuhan University.

Leave a Comment or Question

Help us make this conference the best it can be for you. Have questions you'd like this speaker to address? Suggestions for issues that deserve extra attention? Feedback that you'd like to share with the speaker and other attendees?

Join the conversation here (requires login)

Comments

Picture of Henry Zeng
Henry Zeng | PRINCIPAL PROGRAM MANAGER ON AI
2019-04-05 02:18 CST

Thanks Yunyi, hope to see you in the tutorial and discuss with you in face over those topics.

Yunyi Zhu | SENIOR DATA AND BUSINESS ANALYST
2019-04-03 21:38 CST

Dear Speakers,

I am interested in how deep learning can help to forecast time series with a hybrid of multiple seasonalities (e.g. yearly, weekly, daily) and external regressors (e.g. holidays, campaigns, forecasted weather). It would be great if you shed some light on this topic. Thank you!